Enhanced Precision in Cell Culture Analytics: Leveraging Artificial Intelligence for Unbiased and Non-Destructive Assessment of Cell Growth and Viability.

Enhanced Precision in Cell Culture Analytics: Leveraging Artificial Intelligence for Unbiased and Non-Destructive Assessment of Cell Growth and Viability. | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Enhanced Precision in Cell Culture Analytics: Leveraging Artificial Intelligence for Unbiased and Non-Destructive Assessment of Cell Growth and Viability. Cheung Pang Wong, Nasrin Khazamipour, Soroush Aalibagi, Louise Ramos, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6280571/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Precise assessment of cell growth, count, and viability is crucial in biological and medical research. Traditional cell analytics involve manual processes, such as cell counting or reagent-based approaches that are user-dependent and prone to bias. Semi-automated systems for counting cells, tracking cell growth, and determining viability, have been introduced over the past decades. However, these methods are often time-consuming, require labeling steps, and involve costly instrumentation and consumables. Changes in cell growth and/or viability create biological patterns that can be interpreted by artificial intelligence (AI). Here, we report the development and validation of SnapCyte™, an AI application that performs accurate, unbiased, label- and reagent-free cell analyses from basic cell culture images. Using cell lines with diverse morphologies in various culture conditions, we generated a comprehensive and fully annotated image database that was used for AI education. Convolutional neural networks were employed for cell localization and iterative training loops until a stable performance of > 95% accuracy was obtained for all readouts. The fully trained AI demonstrated high Precision and Recall and performed with greater accuracy and less variation as compared to standard methods. As the SnapCyte™ analyses are performed on cell images only, data acquisition is non-invasive to the experimental setup, enabling real-time use of cells in downstream assays. In summary, SnapCyte™ is a fast and accurate cell analytics platform, resistant to user variations and independent of reagents or specific equipment, with improved performance over current cell analytics methodologies. Biological sciences/Biotechnology Biological sciences/Cell biology Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 INTRODUCTION Routine quantification of confluency, cell counts, and cell viability are critical steps in life science research and biopharmaceutical industry workflows [ 1 – 3 ]. Accurate and consistent measurements of these parameters are essential for data reproducibility and experimental success. Confluency refers to the percentage of a surface area covered by a layer of cells in a culture vessel [ 2 ]. Regular determination of confluency is commonly used as a readout for cell growth and fitness in research fields such as cancer biology, stem cell research, and regenerative medicine [ 4 , 5 ]. In any experimental setup, maintaining optimal cell density is crucial, as high confluency can affect growth dynamics, delay cell passaging and harvesting, and impact downstream experiments. Overcrowding can limit nutrient access, reduce cell viability, and cause cell detachment from the culture plate surface [ 6 ]. Conversely, low confluency can result in insufficient cell-to-cell contact, impairing cell signaling and growth [ 6 , 7 ]. Growth dynamics and cell concentrations depend on inoculum cell count [ 7 , 8 ]. Cell counting is used for determining seeding density but also for tracking cell growth rates particularly in non-adherent cell cultures, providing information on cellular fitness and/or responses to experimental treatments. Obtaining precise cell numbers is required for seeding cells in culture plates to be used in subsequent experimental setups, normalizing results, and optimizing experimental conditions for procedures such as cell transfections [ 9 , 10 ]. In clinical settings, precise cell counts are crucial for determining healthy ranges of cell populations, assessing toxicity, or evaluating immune reactions and other clinical parameters. Accurate cell counts are therefore vital for both life science research and clinical outputs [ 11 , 12 ]. Tracking cell growth is instrumental for evaluating fitness of cell cultures and the proliferation of live cells serves as an important indicator of culture conditions. Changes in growth rates can provide information on cellular responses to stimuli or drug treatments. Given its importance, there are various methods for measuring cell growth. They can roughly be categorized into manual, colorimetric, and image-based methods. Examples of these methods include chemical dyes ( e.g. , MTT (3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide) assay), manual enumeration [ 13 ], automated cell counters [ 14 ], manual image analyses ( e.g. , Image J) [ 15 ], and automated image processing methods ( e.g. , Incucyte®) [ 16 ]. While these techniques are valuable, the readouts are affected by human bias and variability in experimental conditions. Conventional determination of cell counts and cytometric parameters, most prominently, alive and intact dead cell ratios are performed using a hemocytometer ( e.g. , Bürker-Türk counting chamber) [ 17 , 18 ]. The major advantage of this conventional approach is the direct observation of the cell culture by the operator, which enables rapid detection of problems such as contamination or aggregation of cells. However, this method is prone to human error that can occur during many stages of the cell enumeration process, including mixing, handling, and dilution of the cells. Biggs et al. performed a study to quantify the standard error among trained technicians who counted red blood cells from one sample of blood by hemocytometers [ 19 ]. They reported that the standard errors between the results of individual experiments conducted by two technicians and five technicians were 3.6% and 7.6% respectively, suggesting that the magnitude of errors increases with the number of individuals in an experiment. Different researchers might interpret mammalian cell boundaries in a hemocytometer differently, impacting results. A more recent study conducted by Manzini et al. showed that the variation can reach nearly 20% among different operators who are highly experienced in cell counting [ 20 ]. Another report documented the effects of dilution factors on total cell concentration measurements from manual methods. The coefficients of variation surged from 0.072 to 0.2 between 0.3 and 0.5 dilution fraction in an experiment, indicating the degree of human error in diluting samples during the cell counting process [ 21 ]. Additionally, the reproducibility of manual cell counts can be low, especially if high cell density cultures are used [ 22 ]. Beyond the issue of human error, manual counting of cells is also a time-consuming task, especially when having to perform multiple counts on the same culture for increased accuracy. This can lead to process fatigue and cause errors due to subjective interpretation of borderline cases. Colorimetric assays, such as MTT (3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide) and crystal violet, are widely used for assessing cell proliferation due to their extensive validation in numerous publications. However, these methods have notable limitations. The MTT assay measures metabolic activity by relying on the reduction of MTT by mitochondrial dehydrogenases to form an insoluble formazan product. This method is influenced by factors such as mitochondrial activity, which may not directly correlate with cell number or viability, leading to inaccuracies. Also, compromised metabolic activity in viable cells can yield misleading results [ 13 ]. Crystal violet staining, which involves binding of dye to DNA in viable cells, is semi-quantitative and influenced by cell size and DNA content variations. It does not distinguish between live and dead cells, potentially overestimating proliferation [ 23 ]. Both MTT and crystal violet assays are destructive, requiring cell lysis that prevents further analysis of the same cell population. Additionally, the definitive endpoint of the assay precludes real-time monitoring of growth and viability, calling for multiple parallel setups that increase the risk of error. The use of plate readers adds to the complexity and cost. Other reagent-based assays, such as WST-1 and CyQuant, offer some advantages but not without notable limitations. The WST-1 assay measures the conversion of a tetrazolium salt to a water-soluble formazan product, which is a more stable reaction compared to MTT. However, the assay still depends on cellular metabolic activity, which can vary independently of cell number [ 24 ]. CyQuant assays use a fluorescent dye that binds nucleic acids providing a more direct and sensitive measurement of cell number. However, CyQuant requires cell lysis, limiting subsequent live-cell analyses [ 25 ]. Collectively, reagent-based methods for assessing cytostasis or cytotoxicity are often static or indirect, evaluating treatment effects at single time points or on associated cellular processes, such as membrane integrity. These methods cannot capture dynamic changes over time and they rely on endpoint measurements that do not reflect temporal variations in cell growth or viability. These limitations highlight the need for more accurate, dynamic, and non-destructive methods for assessing cell growth and viability in cell cultures. Image-based methods measuring cell confluency (instrument-linked methods such as IncuCyte®) have gained popularity in recent years. While offering the convenience of automation and real-time tracking, these instruments are costly, requiring consumables and maintenance, and have limited user capabilities, making them inaccessible for many laboratories. Other image-based methodologies have also been developed using image processing, or more recently, machine learning (ML) to overcome user variability and reduce processing time, but they come with their own limitations. For example, the method developed by Soleimani et al. involves a multi-step process with image normalization, contrast enhancement, denoising, binary image conversion, modification, and finally measurement [ 26 ]. This method's empirical modification step, akin to morphological opening [ 27 ], may lead to incorrect cell and noise classification, as manual adjustments in image processing can introduce bias and misclassification. Other methods struggle with extreme density scenarios. Wang et al. reported low accuracy (F-score < 80%) for low confluency images (< 40%) and the method relied on artificial rules, such as thresholds, which hindered algorithm generalizability [ 28 ]. Furthermore, the deep learning-based approach presented by Ayanzadeh et al. [ 29 ] shows great potential but focuses solely on low confluency cases, neglecting high confluency scenarios, thereby highlighting the need for improved methods. Overall, while current image-based methods offer advancements over reagent-based techniques, they require further development to address limitations and improve accuracy and accessibility for a wider range of laboratories. Transitioning from these traditional methods, deep learning-based computer vision emerges as a superior approach, setting new benchmarks in image-based cell analytics. Although there have been attempts to integrate various ML technologies into cell analysis, these methods require further enhancement to accommodate precise detection of clumped or overlapping cells, and adaptability to different cell morphologies and imaging conditions [ 30 – 32 ]. For instance, the methodology developed by Jiang et al. employs a random forest and density map framework utilizing handcrafted features that limit its performance due to a limited receptive field [ 33 ]. Similarly, the approach by Schmidt et al. uses a density-regression deep convolutional neural network, combining a U-Net-like density predictor with a VGG-like regression model [ 31 ]. Although innovative, this approach struggles with detecting clumped and overlapping cells. Cellpose offers a robust segmentation algorithm that performs well across various cell types with minimal adjustments [ 32 ]. However, this method falters under varying imaging conditions and morphological changes in our experiments, affecting its accuracy and limiting its general applicability to different data types. These limitations underscore the need for more robust and objective methods for cell growth quantification. Here, we describe the development and performance of a novel AI platform technology for advanced cell analytics that can be easily integrated with most laboratory workflows. RESULTS Deep Learning Optimization for Cell Confluency Detection Cell growth patterns are reflected in numbers and densities of cells in a defined area or volume. These patterns can be captured in microscope images that are also particularly well suited for AI assessment. To build AI capacity able to assess growth patterns of cells, we designed a simple workflow that utilized an iterative human-in-the-loop approach for fine-tuning the confluency.Starting with a naive U-Net model, multiple training cycles using human annotated and corrected sets were needed to ultimately reach output performances of > 90% accuracy (Fig. 1A). We first determined an AI model that could adequately report on cell density or confluency as a readout for cell growth. We employed a convolution-based encoder-decoder architecture based on U-Net for training of the AI (Fig. 1B) . The U-Net architecture, characterized by its "U" shape, includes a contracting path, a bottleneck, and an expansive path, totaling 23 convolutional layers. Each segment of the architecture is critical for capturing features at different resolutions and effectively combining them for precise pixel-wise classification [ 34 ]. We utilized Accuracy, Precision, Recall and F1-score parameters to assess the pixel-level segmentation performance of our model quantitatively. The Accuracy parameter provides a comprehensive measure of segmentation performance on both foreground and background [ 35 ], while the F1-score focuses on the accuracy of target segmentation by the network [ 36 ]. Following four training iteration cycles, the model achieved a desirable threshold of Accuracy (0.94), and F1 (0.91) (Fig. 1C) . Robustness of Confluency Detection Across Heterogeneous Imaging Conditions and Cell Types Since the SnapCyte™ ML model is designed to be user-independent, with no need for users to set parameters, its universality is crucial. This means the model must perform effectively with images acquired in diverse settings and across a wide range of cell types. In the first step, we compiled and annotated a test set of 86 images, covering a broad spectrum of image qualities and sizes. This test set consisted of 43 published images of ATCC’s top 23 cell lines, available at both low and high confluency, and 43 images from 42 cell lines, as shown in 18 publications [ 37 – 53 ]. SnapCyte™ analyzed all images, generating masks that were compared to HQP annotation masks ( Fig. 2A and Tables S1, S2 ). SnapCyte™ achieved a strong correlation with manual measurements (𝑅2 = 0.98) and a slope of 0.96, reflecting near-perfect alignment. We further tested SnapCyte™ against existing ML models using the ATCC image dataset (𝑛= 41), including a modified Cellpose model [ 32 ], the approach described in Lobsenz et.al [ 54 ], and a widely accessible community-contributed method “Otsu-Based method” (ht tps://github.com/insilicomab/cell_confluency) . Our analyses demonstrated a strong correlation between manually assessed confluency values and those generated by SnapCyte™ at both high and low cell density conditions, with an R² of 0.98. In contrast, the Cellpose, Lobsenz, and GitHub models showed weaker correlations, with R² values of 0.59, 0.42, and 0.43, respectively (Fig. 2B) . We also assessed the error metrics where the Mean Absolute Error (MAE) of SnapCyte™ was significantly lower at 2.8, indicating exceptional accuracy with minimal deviation from actual values. Conversely, the other models displayed considerably higher MAEs of 15.4, 20.5, and 36, suggesting less precise predictive capabilities and significant average errors. The Root Mean Squared Error (RMSE) for SnapCyte™ was also low at 3.7, highlighting the model's accuracy and error consistency. The other models, such as GitHub, Lobsenz, and Cellpose, recorded RMSEs of 24.9, 28.4, and 40.8 respectively, indicating moderate to significant errors that could compromise model reliability and effectiveness in practical applications (Fig. 2C) . Furthermore, we explored the model's robustness concerning image resolution by analyzing 10 images of diverse cell lines (e.g., MCF7, MG63, LNCaP, VCaP, IGRCaP-1, NCIH660) at 80%, 50%, 20%, and 8% of their original resolutions. The confluency value deviations between these resolutios were less than 2% (Fig. 2D) , showcasing the model's robustness against variations in image quality. Overall, our data demonstrate that SnapCyte™ maintains high robustness, universality, and independence from experimental settings such as microscope lighting or image resolution. Optimizing Image Sampling for Efficient and Accurate Confluency Estimation in Cell Culture Vessels at Various Magnifications Since SnapCyte™ relies on sample images to predict total vessel density, determining the minimal number of images needed for accurate measurements is essential. We investigated the time efficiency of capturing images across various cell culture vessels at different magnification levels to accurately estimate vessel confluency. At 4,000× and 10,000× magnification, the minimum number of images required to achieve a standard deviation of less than 5% for 96-well plates, 12-well plates, 6-well plates, and 10 cm dishes was found to be 2, 2, 4, and 5, respectively (Fig. 3A, 3B) . This corresponds to 36.5%, 3.0%, 0.6%, and 0.3% of the total vessel area at 10,000× magnification, and 22.8%, 18.9%, 13.8%, and 2.1% at 4,000× magnification, respectively. The total imaging times for these vessels (all wells in a vessel) were 32, 4, 4, and 1 minute, respectively. However, for 96-well plates at 4,000× magnification, a single image covering 11.4% of the well was sufficient to accurately predict the confluency of the entire well (Fig. 3B , upper panel ), reducing the total imaging time to approximately 16 minutes for the whole vessel. Although images were acquired manually, the low number of required images per vessel and the rapid analysis by the ML model make SnapCyte™ an efficient method for assessing confluency or density of cells in most basic laboratory settings. This efficiency could potentially be further enhanced by integration with an automated imaging system, offering even faster and more consistent results. Benchmarking SnapCyte™ Confluency Detection Against Standard Proliferation Assays Accurate and efficient monitoring of cell growth is essential for biological research and drug development studies. Traditional cell proliferation assays either track the number of cells over time or rely on surrogate readouts, such as metabolic activity, ATP or the percentage of area covered by adherent cells. We aimed to investigate the correlation between confluency measurements by SnapCyte™ and cell counts. We first plated increasing concentrations of various cell lines, including epithelial (MCF7, PC3), mesenchymal (MG63), and fibrosarcoma (HT1080) cells, and assessed cell confluency 6 hours after plating using SnapCyte™. The correlation between cell counts and confluency was evaluated across three independent experiments. Results demonstrated a strong correlation between confluency assessed by SnapCyte™ and cell count, with a coefficient of determination (R²) greater than 0.9412 (Fig. 4A; Fig. S1 A, C, E, G). Next, we performed a similar analysis 24 hours after plating where wells seeded with 1.00E + 06 cells had reached near-complete confluency (~ 100%). The data showed a high linear correlation (Fig. S1 B, D, F, H). Linear regression analyses across all experiments, including four cell lines with three replicates at each time point, confirmed a robust correlation (average R² = 0.974) (Fig. 4B). Additionally, analysis of linear regression equations from replicates in each experiment revealed consistent slopes and interception points, underscoring the high reproducibility of the correlation ( Fig. S1 A-H ). Collectively, these findings demonstrate that SnapCyte™ can use cell confluency to precisely report on cell numbers and proliferation rates in 2D cell cultures across diverse cell lines. To further establish the performance and reliability of SnapCyte™ as a tool for precise monitoring of cell proliferation, we tested its accuracy and correlation in measuring cell growth across various techniques and assessed its performance in terms of ease of use, time efficiency, and reproducibility. We benchmarked SnapCyte™ against established reagent-based proliferation assays, including Crystal Violet (CV) staining, CyQuant, and WST-1. Initially, we evaluated the correlation between the normalized viability readouts from each assay and the initial cell counts at time of seeding. We found a high correlation across all techniques with no significant statistical difference in the slopes between SnapCyte™ and the three alternative assays. Similarly, no significant difference was observed in SnapCyte™ data collected by two independent users (F = 0.53, dfn = 4, dfd = 35, P = 0.71) ( Fig. 4C ). We also assessed the variability (delta) of normalized viability readouts among the different methods. The delta between SnapCyte™ users was < 0.5%, as compared to 2% for the colorimetric CV assay and 4.5% for the metabolic assays (WST-1 and CyQuant) ( Fig. 4D ). In a subsequent experiment, we employed SnapCyte™ to monitor cell growth (measured as confluency over time) and compared these data against the conventional CV method ( Fig. 4E ). The data were similar with a delta value of 1.8% between the averages of the two methods ( Fig. 4F ). Additionally, no significant difference was observed between the two methods at any of the time points. To compare, SnapCyte™ provided fast (15 min/time point), accurate, and time-lapsed growth data, facilitating direct use of the cell cultures in downstream applications. By contrast, the CV method necessitated multiple cultures for endpoint readouts and required more than 2 hours of processing time. We next evaluated the performance of SnapCyte™ as a readout for cytotoxicity relative to the live cell imaging system IncuCyte® and the MTT assay. In this study, we assessed the cytotoxic effects of docetaxel on PC3 cells at two different concentrations. The experiments were conducted in triplicates and repeated by independent users. Results demonstrated that SnapCyte™ generated comparable values to both IncuCyte® and the MTT assay, with SnapCyte™ and IncuCyte® displaying lower standard deviations than the MTT assay (Fig. 4G ). Additionally, SnapCyte™ exhibited less variability between independent experiments, further emphasizing its reliability and output reproducibility. To measure inter-user reproducibility, the same biological condition was plated four times in nine replicates and measured by four experienced users using IncuCyte®, SnapCyte™, and MTT. The data indicated no difference between techniques and users with SnapCyte™ and IncuCyte® (standard deviation 14% and 13%, respectively), and showed that these methods achieved less variability as compared to the MTT assay (27%) ( Fig. 4H ). This underscores the advantages of non-invasive measurement methods in providing consistent, user-independent data. Moreover, when the same experiment was measured by four different users setting their own parameters, paired measurements showed an average standard deviation of 6.5% and 7.5% for SnapCyte™ and IncuCyte®, respectively. To validate the performance of SnapCyte™ in drug testing applications, we evaluated the half maximal inhibitory concentration (IC50) of docetaxel across three distinct cell lines: MCF-7, PC3, and IGR-CaP1. Each cell line was tested in three independent experimental setups, and the IC50 values were calculated using transformation and sigmoidal regression analyses. SnapCyte™ demonstrated robust correlation coefficients for PC3 (0.935), MCF-7 (0.982), and IGR-CaP1 (0.972), with IC50 confidence intervals of [0.9–1.5], [0.59–0.75], and [1.5-2], respectively. These values were compared to those obtained with IncuCyte® and the MTT assay, which generated lower correlation coefficients and wider IC50 confidence intervals of 0.707 [0.7–2.7], 0.919 [0.68–1.13], 0.843 [1.4–3.1] for IncuCyte® and 0.649 [0.7-3], 0.761 [0.24–0.74], 0.865 [1.35–2.85] for the MTT assay in the respective cell lines (Fig. 4I, J). Lastly, we evaluated the SnapCyte™ model's ability to replicate published data by determining the doubling time of three cell lines across three independent experiments. Our results not only aligned closely with each cell line’s published doubling time, but also showed that the doubling time values were near the median of values from independent studies ( Fig. 4K, L ). Combined, the results indicate that SnapCyte™ provides highly accurate and reproducible measurements of cell growth and confluency across diverse cell lines. The system’s non-invasive and time-lapsed capabilities make it an efficient and reliable alternative to conventional instrument and reagent-based assays. SnapCyte™ offers comparable accuracy, reduced variability, and faster processing times and is well-suited for routine growth and survival assessments in cell biology laboratories. Optimization of Deep Learning Models for Cell Counting and Viability Assessment Accurate cell counts are essential for cell biology assays, particularly for non-adherent cells where confluency measurements are inadequate. We next aimed to optimize an ML model for assessing cell numbers, employing cellpose model [ 32 ]a standard U-Net architecture pre-trained on a diverse range of cellular images (Fig. 5A) . We used various cell lines, beads, PBMCs, and RBCs that were loaded onto a hemocytometer to acquire images for annotation. The ML model utilized a combination of down-sampling and up-sampling techniques to process feature maps. Training and fine-tuning the Cell Count model using our human-in-the-loop approach (Fig. 1A) required seven iterative cycles to achieve Precision and Recall rates of over 95% in detecting cells while excluding debris (Fig. 5B) . However, further improvements posed challenges due to a 5% variation in labeling, which may result from discrepancies even among cell annotation experts. A drop in the F1 score during cycle 4 was attributed to the introduction of a more complex and variable dataset, reflecting more realistic scenarios. After cycle 6, our analysis indicated that annotation inconsistencies were the primary obstacle for further improvements, as supported by previous research by Kang et al. [ 55 ]. A dataset review identified and resolved these inconsistencies, leading to improved performance in the final cycle, achieving 95% Precision and Recall at cycle 7. We further evaluated the performance of SnapCyte™ in segmenting and counting single cells against other state-of-the-art models (Cyto1 [ 32 ], Cyto3 [ 56 ], Omnipose [ 57 ], and StarDis [ 31 ] on a randomly selected dataset of 50 images. The results reveal that SnapCyte™ consistently outperforms the alternatives across all three key metrics: precision, recall, and F1-score (Fig. 5C) . SnapCyte™ achieves the highest average precision of 94%, significantly surpassing Cyto1 and Cyto3 (both at approximately 75%), and far exceeding Omnipose (37%) and StarDist (48%). Similarly, the recall of SnapCyte™ is 97%, considerably higher than Cyto1 (72%) and Cyto3 (78%), with Omnipose (63%) and StarDist (42%) demonstrating inferior performance. Finally, SnapCyte™ achieves the highest F1-score at 95%, reflecting its superior balance between precision and recall, compared to Cyto1 (73%), Cyto3 (74%), Omnipose (49%), and StarDist (45%). Among the tested models, Cyto1 and Cyto3 demonstrate moderate and consistent performance across all metrics, though they lag behind SnapCyte™. Omnipose exhibits substantial deficiencies in precision and F1-score, likely due to its reduced ability to accurately detect individual cells. StarDist shows poor overall performance in this diverse data set, with significantly low precision and recall, indicating challenges in both cell detection and segmentation. These results highlight the robustness and accuracy of SnapCyte™, establishing SnapCyte™ as the most reliable tool for cell counting among the models tested. To assess model performance across various cell densities, we evaluated Precision, Recall, and F1 scores using subsets of our dataset compiled based on annotated cell numbers per image. We analyzed 12 images with fewer than 50 cells, 12 images with 50 to 300 cells, and 12 images with more than 300 cells. The data showed consistent and strong performance across all groups (Fig. 5D) . As the number of cells increased, the metrics (Precision, Recall, and F1) improved, reaching 0.95 for images with more than 50 cells. Although we did not determine an upper limit for cell numbers, the model’s robustness in denser images suggests its suitability for applications involving cell clumping and overlap. Given the importance of precise cell size measurements in biological assays, particularly for distinguishing cell types, we optimized a Size Estimation model, which works in conjunction with the Cell Count model. The process involves two steps: first, the Cell Count model evaluates images using a default diameter to compute a Style array—256 float values array, representing image features. A linear regression model then predicts cell size based on the aforementioned Style array. Finally, the Cell Count model generates output-masks using the median of the cell sizes, as determined by the Size Estimation model. The image is resized according to the predicted diameter, and the Cell Count model generates output masks, from which the median object size is determined as the final predicted size. Utilizing a visual representation where the x-axis represents the true cell diameter and the y-axis the predicted cell diameter (in pixels), we observed a close alignment between predicted and actual sizes (R² = 0.95) ( Fig. 5E ). This strong correlation confirms the accuracy of the Size Estimation model, making it a reliable tool for downstream applications requiring precise cell size measurements. The combination of robust predictions from the Cell Count model and linear regression ensures accurate size estimations with low error rates suitable for most cell counting requirements. In many cell biology assays, cell count is meaningful only when combined with viability assessments. Thus, we aimed to develop an ML model capable of distinguishing live and dead cells based on Trypan Blue or Erythrosin-B staining, two commonly used dyes for cell viability assessment [ 58 ]. To determine cell viability, images of cells stained with these dyes first undergo analysis by the Cell Count model to identify individual cells. The resulting segmentation mask, combined with the original image, is then processed through a separate U-Net network. This U-Net architecture comprises four convolutional layers with ReLU activations and max pooling for feature extraction, followed by three transposed convolutional layers for spatial restoration. The final layer employs a convolutional operation with a Tanh activation function to generate a single-channel segmentation mask for accurate viability detection. The model was trained on a dataset of 244 images, with an additional 244 augmented images to enhance robustness. Our validation and test set include 62 and 77 original images respectively. Evaluation metrics, including Accuracy, Precision, Recall, and F1-score, are depicted in (Fig. 5F ), alongside an example output image demonstrating the accuracy of the viability detection. Among the available machine learning models for cell counting and segmentation, to our knowledge, only the model published by Kuijpers et al [ 18 ] was reported to distinguish live and dead cells using trypan blue staining. Using a dataset of 20 images of trypan blue-stained cells loaded into different types of chambers (e.g., hemocytometers and KOVA slides), we compared SnapCyte™ to the Kuijpers model for total cell count and live/dead cell discrimination. For total cell count, Kuijpers exhibited high error rates (MAE: 57.4%, MSE: 3656, RMSE: 60.5%), indicating consistently large errors, with occasional extreme outliers contributing to the higher RMSE. In contrast, SnapCyte™ achieved significantly lower errors (MAE: 2.4%, MSE: 26.7, RMSE: 5.2%), demonstrating both accuracy and consistency. Similarly, for live and dead cell classification, SnapCyte™ maintained minimal errors (MAE: 2.3%-2.6%, RMSE: 3.1%-4.2%), compared to Kuijpers' higher errors (MAE: 34.2%-23.8%, RMSE: 42.6%-28.4%) ( Fig. 5G ). Although the Kuijpers model had false positive and false negative rates below 2% for live/dead classification, its higher overall errors stemmed from poor cell segmentation, highlighting the importance of accurate segmentation. The low and closely aligned MAE and RMSE across all tasks reflect precise and reliable predictions and confirm robustness and superior performance of SnapCyte™ in cell analyses. Accurate detection and counting of particles of different sizes is critical when working with mixed populations of cells or particles, such as those commonly used in co-cultures, primary cultures, and patient samples. To ensure that our model could handle diverse sample types, we tested its capability to count shapes of different sizes. We used 6µm, 10µm, and 16µm flow cytometer size reference beads, which were diluted to different concentrations or mixed in a 1:1:1 ratio. The number of beads in each sample was compared between SnapCyte™ and manual counting. Our results showed a strong correlation between the SnapCyte™ values and expected concentrations for all three bead sizes, both separately and when mixed together ( Fig. 5H-K ). Specifically, the coefficient of determination for the 16µm beads was as high as 0.9999 and > 0.99 for 6µm and 10µm beads. This indicates that our model can accurately detect and count homogeneous and heterogeneous samples with different particle sizes, highlighting its robustness in handling mixed cell populations ( Fig. 5K ). Comparative Performance of the SnapCyte™ Cell Count Model and Conventional Cell Counting Methods We first tested the SnapCyte™ size estimation model using a diverse set of biologically relevant samples. A dataset of 96 images of human red blood cells (RBCs), human peripheral blood mononuclear cells (PBMCs), PC3, and MCF7 cells, was employed. The predicted values of SnapCyte™ were comparable to absolute cell sizes determined by scientists (Fig. 6A) . The results demonstrated a strong correlation between SnapCyte™ and the actual cell sizes for all three types of samples (R² = 0.9733) (Fig. 6B) . Furthermore, the Accuracy, Recall, and F1-score of the model were consistently above 90%, regardless of the cell type analyzed (Fig. 6C) . These findings highlight the ability of SnapCyte™ to accurately determine and differentiate cell sizes across various sample types, including heterogeneous PBMC samples. Next, we benchmarked the SnapCyte™ model’s counting and viability capabilities against manual assessments performed by experienced scientists. We used a test dataset of 128 images featuring Trypan Blue- or Erythrosin B-stained PC3 and MCF7 cells with varying viability percentages. Cells were loaded onto hemocytometers or KOVA slides for evaluation by the SnapCyte™ viability module. Results showed that the average absolute difference between SnapCyte™ and manual counting was less than 5% for both Trypan Blue and Erythrosin B staining across different live/dead cell ratios (Fig. 6D, 6E ). Importantly, the model’s performance was independent of the slide type when tested on KOVA slides ( Fig. S2 ). This demonstrates that the SnapCyte™ viability model is a reliable tool for quantitative microscopy, providing accurate viability measurements for cells stained with two commonly used viability dyes. Most automated cell counters have a reliable detection range of 1.00E + 05 to ~ 1.00E + 07 cells/mL [ 59 ]. This range limitation sometimes requires cells to be concentrated or diluted to obtain accurate counts. We tested the SnapCyte™ model’s performance across a wide range of cell concentrations to establish the dynamic range. Data showed that SnapCyte™ provided reliable cell counting in the range of 1.00E + 04 cells/mL to 2.50E + 07 cells/mL (Fig. 6F ), as indicated by a high coefficient of determination (R² = 0.9847), comparable to Bio-Rad (R² = 0.9953) and manual counting (R² = 0.9978). We further assessed the ability of SnapCyte™ to count cells in serially diluted PC3 cell samples. High-resolution images revealed that each cell was effectively segmented even in the highest concentration of 2.50E + 07 cells/mL ( Fig. 6G ). This demonstrates that SnapCyte™ accurately distinguishes individual cells in densely packed samples that are difficult to count in manual settings. Finally, we evaluated the reproducibility of SnapCyte™ compared to manual counting. Four researchers manually counted MCF7 cells at three concentrations (5E + 06 cells/mL, 1.6E + 06 cells/mL, and 2.9E + 06 cells/mL) using a hemocytometer. In parallel, each researcher determined the cell count using SnapCyte™. As a reference, cell counts were also measured using the TC-20 cell counter. SnapCyte™ results showed no statistically significant differences between SnapCyte™ and manual counting measurements. It also showed smaller standard error deviations for SnapCyte™ measurements compared to manual counting (Fig. 6H ). All together our data demonstrate that SnapCyte™ provides a robust, accurate, and efficient ML-based solution for assessing cell confluency, counting, and viability across diverse cell types and experimental conditions. Its U-Net architecture achieves > 90% accuracy for confluency detection, with strong reproducibility and minimal deviations under varying image qualities and densities. SnapCyte™ also demonstrated high correlation with traditional proliferation assays and outperformed manual counting methods, extending its dynamic range to 2.50E + 07 cells/mL. The system’s non-invasive, time-efficient, and highly reproducible measurements, make it a reliable alternative to conventional assays, offering significant advantages for routine cell biology workflows and drug development studies. DISCUSSION Scientists face numerous challenges related to basic cell analytics proficiency. Research costs have increased drastically, the demand for faster methods have intensified, and the scientific community has encountered a reproducibility and replicability crisis [ 60 ], which reflects a multifactorial phenomenon where standardization and accuracy of technologies play significant roles. In cell analytics, traditional methods, such as manual cell counting and reagent-based approaches, are time-consuming, prone to human error, and require costly instrumentation. Additionally, recent image processing technologies have demonstrated limitations, including being time-consuming and requiring users to define subjective parameters ( e.g. , ImageJ) or they are inaccessible to many research groups due to cost. There is a critical need in cell biology research to eliminate user variability inherent in conventional methods and to develop accessible and reliable solutions to ensure data comparability between studies. AI-based systems have emerged as transformative tools in cell analytics, promising rapid, unbiased, and consistent cell analysis readouts. Recent advancements in automation and ML for basic cell culture tasks have shown promise, but only a few have proven to be accurate and applicable to a wide range of cells and experimental conditions. Many ML models currently available are trained on open-source datasets that feature idealized, clean images, often lacking the variability encountered in everyday cell culture environments. These datasets typically do not reflect the real-world challenges, such as debris, clumps, non-uniform lighting, or irregularities in sample preparation, making their models less effective in practical applications. Consequently, there is still an imminent demand for accurate and affordable deep learning models to be widely deployed in research. SnapCyte™ was developed using a rigorously curated dataset encompassing a wide array of cell lines and conditions, meticulously annotated to customize AI models for detection and analysis of cells across diverse experimental setups. We have customized three ML models to measure cell confluency, cell count, and cell viability—three readouts that constitute the basics of cell analytics in a cell biology or life science laboratory. SnapCyte™ achieves high precision and recall for all three models across all cell lines and image qualities that were tested, reflected by high F1 scores. Notably, for the Cell Count ML model, accuracy increased in samples containing a higher number of cells, proving the superiority of this method in counting concentrated samples. This is important, as dilution affects the prediction accuracy of total cell counts [ 21 ]. SnapCyte™ consistently outperforms existing publicly available models with exceptionally low error metrics, demonstrating high precision and reliability. While some models achieve a balance of accuracy and consistency, others exhibit significant errors across all metrics, indicating a need for substantial adjustments or alternative modeling approaches to achieve acceptable predictive accuracy. This highlights the importance of customization in machine learning models. Generalist models like Cellpose are designed for broad applications, which can limit their precision, whereas SnapCyte™ was developed using specialized and diverse datasets, ensuring superior accuracy and performance. Confluency measurements are crucial for ensuring the consistency and reliability of cell culture experiments. Traditional methods often rely on subjective visual assessments, leading to significant errors and variability. Studies have shown that confluency eyeballing errors can be very high, especially at higher confluency levels, impacting experimental outcome and reproducibility. For example, a study published by Lin et al., discusses the challenges in accurately determining cell confluency, highlighting that visual estimations can greatly vary among researchers, especially among those with less experience. This variability can exceed 30%, affecting the consistency of experimental results [ 61 ]. High confluency levels can lead to contact inhibition, altering gene expression involved in cell cycle regulation and apoptosis, thus affecting studies on cell growth, migration, and response to treatments. Our validation tests of the confluency model on independent datasets, which included a high number of cell lines and random images from publications, demonstrated the universality of the ML model and its applicability to a wide range of laboratory settings (lighting, image resolution, and cell models). The F1 score remained above 91%, and the average deviation from manual masks was 3%. The SnapCyte™ confluency model is highly relevant for data standardization in cell biology research. It enables scientists to instantly evaluate seeding homogeneity and report it accurately. This is crucial for experiments that are sensitive to seeding variation, such as genome editing and differentiation studies, thereby enhancing replicability and reproducibility. This model also assists scientists in maintaining healthy cell cultures by ensuring that cells are seeded homogeneously and passaged at adequate confluency to prevent genetic drift in cell cultures. Additionally, SnapCyte™ allows for non-invasive, time-lapse cell growth measurement from the same vessel, enabling scientists to determine the optimal timing for treatments or cell editing. This capability helps avoid potential issues related to over or under-seeding, which can significantly affect experimental outcomes. Balancing the positive considerations, it is important to acknowledge the limitations of image-based analysis, particularly when based on cell confluency. In some conditions, treatments or cell manipulations may lead to morphological changes and alterations in cell attachment, which can affect cell size and confound confluency measurements. This could result in cell growth assessment errors. Nevertheless, because the analysis is image-based, scientists can visually inspect cell morphology and detect changes, thereby mitigating the risk of misleading information. Accurate cell counting is equally critical in cell biology research. Manual counting methods are prone to significant errors, while newer automated counters struggle with complex samples containing cells of different sizes. SnapCyte™ addresses these challenges by accurately counting particles of various sizes, as demonstrated in correlation studies with manual counting (r > 0.99). Notably, this model can count heterogeneous samples containing beads of varying sizes (6, 10, and 16 µm), further validated by its ability to distinguish different sizes accurately, including distinct populations of RBCs, PBMCs, and cancer cell lines. Additionally, SnapCyte™ exhibits exceptional accuracy in detecting cells stained with Trypan Blue or Erythrosine B to assess viability. As many laboratories transition away from Trypan Blue staining due to its toxicity, the ability to use Erythrosine B provides a safer alternative. This versatility makes SnapCyte™ an impactful tool for measuring cell viability in complex biological samples. Since SnapCyte™ uses images to analyze cells, an important consideration is how representative the images are of the entire vessel especially in the context of cell proliferation. Like many other techniques in cell culture data analysis, cell confluency is based on image sampling, and in the case of adherent cells, the homogeneity of the culture is a cornerstone for performing accurate and reproducible data analysis. Assuming homogeneous seeding, SnapCyte™ requires a minimal sampling number ( i.e. 1–2 for a 96-well plate, 2–4 for 6- and 12-well plates and a 10 cm vessel). This ensures that SnapCyte™ analyses are fast, and return data within a few minutes if using manual imaging. The ability of SnapCyte™ to process data within minutes without necessitating invasive procedures aligns with the pressing need for versatile analytical tools that are compatible with downstream processes. We also investigated whether the cell confluency readout from SnapCyte™ could be used as a surrogate readout for cell proliferation and cytotoxicity assays. First, the correlation between cell counts and SnapCyte™ confluency was > 0.95 in all conditions tested, indicating that confluency could predict cell number and, hence, proliferation ( Fig. 4A and Fig. S1 ). When compared to reagent-based assays, results were comparable to standard methods with a low delta variation. When compared to Crystal Violet, the most commonly used manual assay for confluency, the variation was lower than 4% in all conditions. This offers the added advantage of avoiding errors related to Crystal Violet staining steps and reducing seeding-related variability between vessels by acquiring data from the same vessel. This approach also reduces the use of toxic reagents, and saves time by allowing immediate use of cells in downstream assays. Compared to the IncuCyte®, data obtained on SnapCyte™ were similarly accurate. Also, data between different SnapCyte™ users were more consistent as parameter settings were predefined. Furthermore, when compared to MTT and IncuCyte®, IC50 values of cell lines treated with docetaxel and calculated with SnapCyte™ were more accurate, as evidenced by a high non-sigmoidal correlation coefficient ( Fig. 4I ). As such, SnapCyte™ produced similar IC50 values as other techniques but with smaller confidence intervals. While SnapCyte™ is not tied to an automation system, manual image acquisition remained within acceptable time limits (15 minutes). SnapCyte™ allows acquisition from any type of vessel and microscope, whereas many other methods and techniques are tied to their own instruments (e.g., IncuCyte®). Moreover, the phone app version of SnapCyte™ transforms a smartphone into a cell analytics platform that can perform instant analyses on the device. Additionally, SnapCyte™ reduces contamination risks by eliminating the need to transfer cells between different rooms for measurement, maintaining the integrity of cell cultures and experimental conditions. Finally, the last validation involved comparing SnapCyte™-generated data with published values of doubling time, demonstrating the accuracy of SnapCyte™ by producing values close to the median and demonstrating highly reproducible results. For the cell count readout, SnapCyte™ uses a hemocytometer but can also be used with any type of slide. It eliminates the need for additional machines and specific slides. SnapCyte™ has a wide dynamic range, allowing quantification of highly concentrated samples and different sizes of cells. The data from SnapCyte™ is user-independent, and the variation is minimal. The validation across a range of cell lines, culture conditions, and analytical parameters attests to the robustness and versatility of the SnapCyte™ system. Future studies should focus on expanding the applicability of SnapCyte™ to other complex biological systems and exploring its integration with other cutting-edge analytical tools. By offering high precision, efficiency, and user-independence, the SnapCyte™ AI tool paves the way for more reproducible, accurate, and insightful scientific research, ultimately accelerating the pace of discovery and innovation in life science research. MATERIALS AND METHODS Model architecture In this study, we evaluated multiple deep learning architectures to determine their suitability for cell detection and segmentation tasks. Models such as U-Net, Mask R-CNN, and YOLO were tested for their ability to identify cellular features in microscopy images. Initial evaluations revealed that U-Net and Mask R-CNN faced challenges in accurately detecting specific features in complex microscopy data. While YOLO showed efficiency in fast object detection, its precision was inadequate for the requirements of life science imaging [ 29 ]. We tested several other models, including those utilized in literature and competition datasets, such as those from Kaggle, which often required specific image types (e.g., grayscale or single-cell fluorescent images). For instance, DeepCell [ 62 ] was designed for single-cell fluorescent microscopy images, while CellProfiler [ 63 ] underperformed on our multi-cell culture images, likely due to its instrument-specific limitations. DeepLab demonstrated promising results in general semantic segmentation tasks, but U-Net emerged as the most effective for cell microscopy, owing to its use of skip connections and its symmetrical structure, which improved feature extraction and boundary delineation [ 64 ]. This architecture proved especially beneficial for estimating confluency levels in complex cellular environments. Hence, we chose U-Net as the core architecture for our cell confluency model. We integrated and fine-tuned the Cellpose model [ 32 ], which was initially pretrained for general cell segmentation. While Cellpose demonstrated robustness and adaptability across various cell types, we identified the need for further fine-tuning to align with our specific application, enhancing precision in cell counting and segmentation under diverse imaging conditions. We performed a comprehensive evaluation of several regression models for cell size estimation, including Support Vector Regressor, Elastic Net Regressor, K-Nearest Neighbors Regressor, Random Forest Regressor, Gradient Boosting Regressor, and Linear Regressor. We conducted experiments using our dataset of images and compared the Mean Squared Error (MSE) across all models, yielding values of 43.5, 40.1, 36.1, 23.6, 12, and 10.9, respectively. Our analysis, based on these experiments, revealed that Linear Regression exhibited greater robustness when applied to unseen images from our dataset, outperforming the other models. Building upon Cellpose's Linear Regression Size Model, we integrated this Linear Regression model into our pipeline for cell size estimation. This addition ensures precise determination of cell diameter size, thereby enhancing the reliability and performance of our cell counting system across a wide range of imaging conditions. In our efforts to develop an optimal solution for detecting cell viability, we initially utilized unsupervised ML clustering algorithms to differentiate between live and dead cells in cell culture images. Unsupervised learning, beneficial in contexts without explicitly labeled data, allows algorithms to independently identify patterns or clusters within the data. However, these algorithms often classify images into two clusters, even when only live or dead cells are present. Among the unsupervised algorithms tested, K-means demonstrated acceptable performance with Trypan Blue-stained images but was less effective with other stains, such as Erythrosin B. Consequently, we decided to employ supervised learning techniques and chose the UNet architecture for viability detection. We conducted further experiments to optimize the UNet model's architecture and parameters, which resulted in satisfactory performance across various types of images. Data Sets Cell images were captured utilizing a microscope adaptor and a variety of cell phones (iOS and Android systems) and microscopes (Nikon ECLIPSE Ts2 inverted microscope, Leica DM1000 LED, Helmut Hund Wilovert AFL 30 Series 8 Inverted Trinocular Microscope) with 4×, 10×, and 20× optical magnification, along with different digital magnifications. For the cell confluency dataset, 1500 images of multiple cell lines cultured in various vessels were captured and annotated. For the cell count dataset, various cell lines (PC3, MCF7, Raji,) and cell samples (RBCs, PBMCs) were cultured, adherent cells were detached, and cells were loaded into a hemocytometer or KOVA slides at different concentrations, both in the presence and absence of Trypan Blue or Erythrosin B. Cell death was induced by heat shock at 80℃ for 15 minutes. The dataset also included images of red blood cells, peripheral blood mononuclear cells (PBMCs) isolated as previously described [ 65 ], and beads of various sizes (6, 8, 10, and 16 µm) ( ThermoFisher; Cat# C16506 and Spherotech Inc; Cat# PPS-6K). Two thousand one hundred images were annotated by experienced scientists to generate cell count and live/dead masks for ML. To enhance the model's generalization capabilities, particularly for out-of-focus images, we implemented a Gaussian Blur technique on a subset of randomly chosen images. Cell lines and cultures MCF-7, PC3, HELA, LNCaP, MG63, and HT1080 cells were procured from ATCC (Manassas, VA, USA). IGR-CaP1 cells were provided by Dr. Chauchereau (Gustave Roussy Institute, France) [ 66 ]. Cells were maintained in their appropriate media supplemented with 10% FBS in a humidified incubator at 37°C with 5% CO 2 . All cells were tested for mycoplasma regularly. IncuCyte® Analysis Cell confluency was measured by IncuCyte® according to the manufacturer's protocol. Briefly, MCF7 cells, PC3 cells, and IGR-CaP1 cells were seeded in triplicate in 96-well plates. Cells were treated with Docetaxel at different concentrations and were then monitored on the IncuCyte® Live Cell Analysis System (Sartorius, USA). Images were taken at 4× magnification and analyzed by IncuCyte® Live-Cell Analysis Systems. SnapCyte™ Count Number and viability of cells were determined by SnapCyte™ Count and SnapCyte™ Viability modules, respectively. Briefly, cells were cultured in various cell culture vessels and cell pellets were collected. 10 µL of each sample was individually loaded on a Hausser Scientific™ Bright-Line™ Phase Hemacytometer (Cat.# 026716) or KOVA™ Glasstic™ Slide (fisher scientific; Cat.# 22-270141) and the cell numbers were determined by SnapCyte™ Count module. For cell viability analyses, equal volumes of Trypan Blue solution, 0.4% (Gibco; Cat.# 15250061) or 0.1% of Erythrosin B (ThermoFisher; Cat.#A14180.14) were added to the resuspended cells in corresponding cell culture media supplemented with 10% FBS before the samples were loaded on a hemocytometer or KOVA™ Glasstic™ Slide (fisher scientific; 22-270141). Cell viability was determined by SnapCyte™. For beads analysis, 10 µL of the Cell Sorting Set-up Beads (for UV lasers) (ThermoFisher; Cat.# C16506) or Polystyrene Particle Size Standard Kit, Flow Cytometry Grade (Spherotech Inc; Cat.# PPS-6K) were prepared at different concentrations with distilled water and loaded on a hemocytometer. Bead numbers were determined by SnapCyte™ Count module. Four photos per sample were taken by an Apple iPhone 8 and an adapter provided by SnapCyte™, mounted on a Nikon ECLIPSE Ts2 inverted microscope at 10× magnification. Region of Interest (ROI) was set up for each photo taken. SnapCyte™ Confluency Confluency of cells was measured by SnapCyte™ Confluency according to the manufacturer's protocol. Briefly, cells were cultured in various cell culture vessels and cell confluency was determined at different timepoints by SnapCyte™ Confluency in photos taken with various phones and SnapCyte™ universal Smartphone Adapter mounted on Nikon ECLIPSE Ts2 inverted microscope. Region of Interest (ROI) was set up for each photo taken and results were exported and analyzed. Microscope magnification used and the numbers of photos taken at different timepoints in experiments are described in corresponding figure legends. Colorimetric assays Colorimetric assays ( i.e. MTT, WST1, Cyquant, and Crystal violet assays) were used to determine the viability of MCF-7, PC3 and/or IGR-Cap1 at 72 hours after treatment of Docetaxel, and subsequently half maximal inhibitory concentration (IC50). Measurements by colorimetric assays were performed according to the manufacturers’ instructions. Briefly, for MTT assay, at 72h after treatment, the media of the plate was aspirated, after which 50 µL of serum-free media and 50 µL of MTT solution (Sigma-Aldrich; Cat.# M2128) dissolved at 5 mg/mL solution in PBS were added into each well. After incubation at 37°C for 3 hours, 150 µL of MTT solvent (4 mM HCl, 0.1% NP40 in isopropanol) was added into each well. The plate was wrapped in foil and put on an orbital shaker for 15 minutes at room temperature. The plate was read by a BioTek microplate reader at OD = 610 nm. Docetaxel IC50 in MTT assay was determined by methods described in the Statistical analysis section below. WST1, CyQuant assay, and Crystal Violet assay were performed as previously described [ 65 , 67 , 68 ]. Statistical analysis The data were assessed using Student's t-test. GraphPad Prism software was used to calculate the statistical significance. The threshold of statistical significance was set at *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001. Determination of IC50 The measured IC50 was calculated using the least square fit of four parameter-sigmoidal curves executed using GraphPad Prism (version 8.4.3) software (GraphPad Software, San Diego, CA, USA, https://www.graphpad.com/scientific-software/prism/ ). Comparison between the measured and predicted values was performed using the correlation coefficient R 2 . Doubling time MCF7 cells, PC3 cells, and HELA cells were seeded in triplicate in 96-well plates at 1.00E + 4 cells/well. Cell confluency was determined by the SnapCyte™ Confluency model at 24h, 48h, 72h and 96h after seeding. Two images were taken per well. Experiments were repeated three times independently and values are expressed in mean. Doubling time (Td) of each cell line in each of the three independent experiments was calculated using the equation below where: Nt is the number of cells at time t; N0 is the number of cells initially at time 0; t is time (days); gr is the growth rate. Number of Cells at Time t (Nt) = N0 * e^(gr * t). Growth Rate (gr) gr = ln(Nt / N0) / t. Doubling Time (Td) = ln(2) / gr. Declarations Competing Interests N. AN., M.D., and N. F. are co-founders of SnapCyte Solutions Inc. S. A. and JM. S. are employed by SnapCyte Solutions Inc., and C. D., D. G., and JM. P. hold shares in SnapCyte Solutions Inc. The authors declare that these affiliations do not influence the objectivity or integrity of the research presented in this manuscript. All other authors declare no competing interests. Author Contribution Conceptualization: N.AN., N.F., C.D., M.D., and D.G.; Methodology: C.P.W., N.AN., N.F., D.G., S.A., N.K.; Investigation: C.P.W., N.AN., S.A., N.F., D.G., N.K., JM.S., and L.R.; Writing – Original Draft: S.A., N.AN., C.P.W., JM.L., and N.F.; Writing – Review & Editing: All authors; Funding Acquisition: M.D., N.F., N.AN.; Resources: M.D., C.D., N.F., N.AN., D.G.; Supervision: N.F., D.G., N.AN., M.D. Acknowledgement We would like to thank Irina Nelepcu for her technical support, and Dr. Anne Chauchereau for providing IGR-CaP1 cells. This work is supported by the Canadian Institutes of Health Research (CIHR-519572) and MITACS (IT37367). Data Availability All data related to the present study are included in the manuscript and supplementary materials. The annotated datasets are protected by University of British Columbia intellectual property and considered trade secrets. Data inquiries can be directed to Dr. Nader Al Nakouzi. References Segeritz, C. P. & Vallier, L. Chap. 9 - Cell Culture: Growing Cells as Model Systems In Vitro , in Basic Science Methods for Clinical Researchers, M. Jalali, F.Y.L. Saldanha, and M. Jalali, Editors. Academic: Boston. 151–172. (2017). Geraghty, R. J. et al. Guidelines for the use of cell lines in biomedical research. Br. J. Cancer . 111 (6), 1021–1046 (2014). Stoddart, M. J. Cell Viability Assays: Introduction, in Mammalian Cell Viability: Methods and Protocols, M.J. Stoddart, Editor. Humana: Totowa, NJ. 1–6. (2011). Freshney, R. I. Basic Principles of Cell Culture , in Culture of Cells for Tissue Engineering . pp. 1–22. (2005). Justus, C. R. et al. In vitro cell migration and invasion assays . J. Vis. Exp. , 2014 (88 ). Rodriguez, E. N. et al. Effect of Seed Cell Density on Specific Growth Rate Using CHO Cells as Model , in Animal Cell Technology: From Target to Market: Proceedings of the 17th ESACT Meeting Tylösand, Sweden, June 10–14 , E. Lindner-Olsson, N. Chatzissavidou, and E. Lüllau, Editors. 2001, Springer Netherlands: Dordrecht. pp. 434–437. (2001). Gregório, A. C. et al. Inoculated Cell Density as a Determinant Factor of the Growth Dynamics and Metastatic Efficiency of a Breast Cancer Murine Model. PLoS One . 11 (11), e0165817 (2016). Ozturk, S. S. & Palsson, B. O. Effect of initial cell density on hybridoma growth, metabolism, and monoclonal antibody production. J. Biotechnol. 16 (3–4), 259–278 (1990). Cocchiararo, I. et al. Back to basics: Optimization of DNA and RNA transfer in muscle cells using recent transfection reagents. Exp. Cell. Res. 421 (2), 113392 (2022). Morten, B. C., Scott, R. J. & Avery-Kiejda, K. A. Comparison of Three Different Methods for Determining Cell Proliferation in Breast Cancer Cell Lines . J. Vis. Exp. , 2016(115). Giannakeas, V. et al. Analysis of Platelet Count and New Cancer Diagnosis Over a 10-Year Period. JAMA Netw. Open. 5 (1), e2141633 (2022). Stella, S. et al. Impact of Different Cell Counting Methods in Molecular Monitoring of Chronic Myeloid Leukemia Patients . Diagnostics (Basel) , 12 (5). (2022). Mosmann, T. Rapid colorimetric assay for cellular growth and survival: application to proliferation and cytotoxicity assays. J. Immunol. Methods . 65 (1–2), 55–63 (1983). Vembadi, A., Menachery, A. & Qasaimeh, M. A. Cell Cytometry: Review and Perspective on Biotechnological Advances. Front. Bioeng. Biotechnol. 7 , 147 (2019). Schroeder, A. B. et al. The ImageJ ecosystem: Open-source software for image visualization, processing, and analysis. Protein Sci. 30 (1), 234–249 (2021). Lanigan, T. M. et al. Real time visualization of cancer cell death, survival and proliferation using fluorochrome-transfected cells in an IncuCyte. J. Biol. Methods . 7 (2), e133 (2020). Davis, J. D. The evolution of the progressive-era hemocytometer. Caduceus 11 (3), 164–183 (1995). Kuijpers, L. et al. Automated cell counting for Trypan blue-stained cell cultures using machine learning. PLoS One . 18 (11), e0291625 (2023). Biggs, R. & Macmillan, R. L. The Error of the Red Cell Count. J. Clin. Pathol. 1 (5), 288–291 (1948). Manzini, P. et al. Validation of an automated cell counting method for cGMP manufacturing of human induced pluripotent stem cells. Biotechnol. Rep. (Amst) . 33 , e00708 (2022). Sarkar, S. et al. Evaluating the quality of a cell counting measurement process via a dilution series experimental design. Cytotherapy 19 (12), 1509–1521 (2017). Zhang, M. et al. Improvement of cell counting method for Neubauer counting chamber. J. Clin. Lab. Anal. 34 (1), e23024 (2020). Feoktistova, M., Geserick, P. & Leverkus, M. Crystal Violet Assay for Determining Viability of Cultured Cells . Cold Spring Harb Protoc. , 2016 (4): p. (2016). pdb.prot087379. Berridge, M. V., Herst, P. M. & Tan, A. S. Tetrazolium dyes as tools in cell biology: new insights into their cellular reduction. Biotechnol. Annu. Rev. 11 , 127–152 (2005). Jones, L. J. et al. Sensitive determination of cell number using the CyQUANT cell proliferation assay. J. Immunol. Methods . 254 (1–2), 85–98 (2001). Soleimani, S., Mirzaei, M. & Toncu, D. C. A new method of SC image processing for confluence estimation. Micron 101 , 206–212 (2017). Sunil, B. et al. Opening and closing in morphological image processing. World J. Adv. Res. Reviews . 14 (3), 687–695 (2022). Wang, Z., Zhan, R. & Hu, Y. Automated Confluence Measurement Method for Mesenchymal Stem Cell from Brightfield Microscopic Images. Microsc. Microanal. 27 (5), 1093–1101 (2021). Ayanzadeh, A. et al. Deep Learning based Segmentation Pipeline for Label-Free Phase-Contrast Microscopy Images . in 2020 28th Signal Processing and Communications Applications Conference (SIU) . (2020). Jiang, N. & Yu, F. A Cell Counting Framework Based on Random Forest and Density Map. Appl. Sci. 10 10.3390/app10238346 (2020). Schmidt, U. et al. Cell Detection with Star-Convex Polygons. in Medical Image Computing and Computer Assisted Intervention – MICCAI 2018. Cham: Springer International Publishing. (2018). Stringer, C. et al. Cellpose: a generalist algorithm for cellular segmentation. Nat. Methods . 18 (1), 100–106 (2021). Jiang, N. & Yu, F. A Cell Counting Framework Based on Random Forest and Density Map. Appl. Sci. 10 (23), 8346 (2020). Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (Springer International Publishing, 2015). Wang, Z. B., Wang, E. & Zhu, Y. Image segmentation evaluation: a survey of methods. Artif. Intell. Rev. 53 (8), 5637–5674 (2020). Wu, B. et al. Investigator accuracy: A center-weighted metric for evaluating the location accuracy of image segments in land cover classification. Int. J. Appl. Earth Obs. Geoinf. 122 , 103402 (2023). Beliveau, A. et al. Tumorigenicity of the mixed spindle-epithelioid SP6.5 and epithelioid TP17 uveal melanoma cell lines is differentially related to alpha5beta1 integrin expression. Invest. Ophthalmol. Vis. Sci. 42 (12), 3058–3065 (2001). Ku, J. L. et al. Establishment and characterization of seven human breast cancer cell lines including two triple-negative cell lines. Int. J. Oncol. 43 (6), 2073–2081 (2013). Gagne, J. P. et al. Comparative proteome analysis of human epithelial ovarian cancer. Proteome Sci. 5 , 16 (2007). Miura, K. et al. Establishment and characterization of new cell lines of anaplastic pancreatic cancer, which is a rare malignancy: OCUP-A1 and OCUP-A2. BMC Cancer . 16 , 268 (2016). Huang, J. et al. The development of radioresistant oral squamous carcinoma cell lines and identification of radiotherapy-related biomarkers. Clin. Transl Oncol. 25 (10), 3006–3020 (2023). Parveez Ahamed, A. A. et al. In vitro antibacterial activity of MGDG-palmitoyl from Oscillatoria acuminata NTAPC05 against extended-spectrum beta-lactamase producers. J. Antibiot. (Tokyo) . 70 (6), 754–762 (2017). Palumbo, P. et al. NOS2 expression in glioma cell lines and glioma primary cell cultures: correlation with neurosphere generation and SOX-2 expression. Oncotarget 8 (15), 25582–25598 (2017). De Vries, G. H. & Boullerne, A. I. Glial cell lines: an overview. Neurochem Res. 35 (12), 1978–2000 (2010). Alemi, M. et al. Anti-inflammatory effect of seeds and callus of Nigella sativa L. extracts on mix glial cells with regard to their thymoquinone content. AAPS PharmSciTech . 14 (1), 160–167 (2013). Holliday, D. L. & Speirs, V. Choosing the right cell line for breast cancer research. Breast Cancer Res. 13 (4), 215 (2011). Zhang, H. Y. et al. Transforming growth factor-beta1-induced epithelial-mesenchymal transition in human esophageal squamous cell carcinoma via the PTEN/PI3K signaling pathway. Oncol. Rep. 32 (5), 2134–2142 (2014). Kiss, A. et al. MTT Test and Time-lapse Microscopy to Evaluate the Antitumor Potential of Nucleoside Analogues. Anticancer Res. 41 (1), 137–149 (2021). Nakamura, R. et al. Establishment and characterization of a human lymphatic endothelial cell line. Biol. Pharm. Bull. 37 (4), 683–687 (2014). Thonhoff, J. R. et al. Compatibility of human fetal neural stem cells with hydrogel biomaterials in vitro. Brain Res. 1187 , 42–51 (2008). Chatree, K. et al. Distinctions in bone matrix nanostructure, composition, and formation between osteoblast-like cells, MG-63, and human mesenchymal stem cells, UE7T-13. Heliyon, 9(5): p. e15556. (2023). Gaberlein, K. et al. Genetic Characterization of Rat Hepatic Stellate Cell Line PAV-1 . Cells , 12 (12). (2023). Shepherd, T. G. et al. Corrigendum: Primary culture of ovarian surface epithelial cells and ascites-derived ovarian cancer cells from patients. Nat. Protoc. 10 (9), 1457 (2015). Lobsenz, A. & Seidler, P. The Confluence Analysis Program: A User-Friendly Tool for Automated Cell Confluence Measurement and Visualization , in Preprints . Preprints. (2024). Kang, C. et al. Variability Matters: Evaluating Inter-Rater Variability in Histopathology for Robust Cell Detection. in Computer Vision – ECCV 2022 Workshops. Cham: Springer Nature Switzerland. (2023). Stringer, C. & Pachitariu, M. Cellpose3: one-click image restoration for improved cellular segmentation. bioRxiv, (2024). Cutler, K. J. et al. Omnipose: a high-precision morphology-independent solution for bacterial cell segmentation. Nat. Methods . 19 (11), 1438–1448 (2022). Kim, S. I. et al. Application of a non-hazardous vital dye for cell counting with automated cell counters. Anal. Biochem. 492 , 8–12 (2016). Seo, D. et al. A Field-Portable Cell Analyzer without a Microscope and Reagents . Sens. (Basel) , 18 (1). (2017). Van Noorden, R. More than 10,000 research papers were retracted in 2023 - a new record. Nature 624 (7992), 479–481 (2023). Chiu, C. H. et al. Systematic Quantification of Cell Confluence in Human Normal Oral Fibroblasts. Appl. Sci. 10 10.3390/app10249146 (2020). Van Valen, D. A. et al. Deep Learning Automates the Quantitative Analysis of Individual Cells in Live-Cell Imaging Experiments. PLoS Comput. Biol. 12 (11), e1005177 (2016). McQuin, C. et al. CellProfiler 3.0: Next-generation image processing for biology. PLoS Biol. 16 (7), e2005970 (2018). Chen, L. C. et al. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE Trans. Pattern Anal. Mach. Intell. 40 (4), 834–848 (2018). Skeltved, N. et al. Bispecific T cell-engager targeting oncofetal chondroitin sulfate induces complete tumor regression and protective immune memory in mice. J. Exp. Clin. Cancer Res. 42 (1), 106 (2023). Chauchereau, A. et al. Stemness markers characterize IGR-CaP1, a new cell line derived from primary epithelial prostate cancer. Exp. Cell. Res. 317 (3), 262–275 (2011). Al Nakouzi, N. et al. Cabazitaxel Remains Active in Patients Progressing After Docetaxel Followed by Novel Androgen Receptor Pathway Targeted Therapies. Eur. Urol. 68 (2), 228–235 (2015). Peacock, J. W. et al. SEMA3C drives cancer growth by transactivating multiple receptor tyrosine kinases via Plexin B1. EMBO Mol. Med. 10 (2), 219–238 (2018). Additional Declarations Competing interest reported. N. AN., M.D., and N. F. are co-founders of SnapCyte Solutions Inc. S. A. and JM. S. are employed by SnapCyte Solutions Inc., and C. D., D. G., and JM. P. hold shares in SnapCyte Solutions Inc. The authors declare that these affiliations do not influence the objectivity or integrity of the research presented in this manuscript. All other authors declare no competing interests. Supplementary Files SupplementalMaterialTable1and2.xlsx Wongetal.Suppfigures.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6280571","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":440054223,"identity":"73ab15dd-54ac-42cd-90a5-dfa4be2dad6a","order_by":0,"name":"Cheung Pang Wong","email":"","orcid":"","institution":"SnapCyte Solutions Inc","correspondingAuthor":false,"prefix":"","firstName":"Cheung","middleName":"Pang","lastName":"Wong","suffix":""},{"id":440054226,"identity":"ed8491fd-55e1-4b58-b396-bf28b81e9e1c","order_by":1,"name":"Nasrin Khazamipour","email":"","orcid":"","institution":"SnapCyte Solutions Inc","correspondingAuthor":false,"prefix":"","firstName":"Nasrin","middleName":"","lastName":"Khazamipour","suffix":""},{"id":440054227,"identity":"11419c61-cc7e-4401-a496-0e89c446d507","order_by":2,"name":"Soroush Aalibagi","email":"","orcid":"","institution":"SnapCyte Solutions Inc","correspondingAuthor":false,"prefix":"","firstName":"Soroush","middleName":"","lastName":"Aalibagi","suffix":""},{"id":440054229,"identity":"4c90d42a-26c2-43db-9183-e4515560f122","order_by":3,"name":"Louise Ramos","email":"","orcid":"","institution":"Vancouver Prostate Centre","correspondingAuthor":false,"prefix":"","firstName":"Louise","middleName":"","lastName":"Ramos","suffix":""},{"id":440054230,"identity":"3a479c84-fa73-4e24-8dac-715029d69df5","order_by":4,"name":"Joya Maria Saade","email":"","orcid":"","institution":"SnapCyte Solutions Inc","correspondingAuthor":false,"prefix":"","firstName":"Joya","middleName":"Maria","lastName":"Saade","suffix":""},{"id":440054231,"identity":"d956770c-1af2-4ef5-a742-c2533169d707","order_by":5,"name":"Casper Dolleris","email":"","orcid":"","institution":"SnapCyte Solutions Inc","correspondingAuthor":false,"prefix":"","firstName":"Casper","middleName":"","lastName":"Dolleris","suffix":""},{"id":440054232,"identity":"e5e76890-b940-4f21-8028-fbd0f7cda210","order_by":6,"name":"Janny Marie L. Peterslund","email":"","orcid":"","institution":"SnapCyte Solutions Inc","correspondingAuthor":false,"prefix":"","firstName":"Janny","middleName":"Marie L.","lastName":"Peterslund","suffix":""},{"id":440054233,"identity":"73d707d2-bf6e-4368-b8df-1b7a5cd3270d","order_by":7,"name":"Daria Golanarian","email":"","orcid":"","institution":"SnapCyte Solutions Inc","correspondingAuthor":false,"prefix":"","firstName":"Daria","middleName":"","lastName":"Golanarian","suffix":""},{"id":440054234,"identity":"0e63edc0-085e-479b-9e2d-4cd83bcbe70f","order_by":8,"name":"Negin Farivar","email":"","orcid":"","institution":"Vancouver Prostate Centre","correspondingAuthor":false,"prefix":"","firstName":"Negin","middleName":"","lastName":"Farivar","suffix":""},{"id":440054235,"identity":"4ec19178-4c8b-4934-88dd-b3f07fbab8f3","order_by":9,"name":"Mads Daugaard","email":"","orcid":"","institution":"Vancouver Prostate Centre","correspondingAuthor":false,"prefix":"","firstName":"Mads","middleName":"","lastName":"Daugaard","suffix":""},{"id":440054236,"identity":"0865f9b9-16cd-428b-9ec4-5a0062d2837f","order_by":10,"name":"Nader Al Nakouzi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABD0lEQVRIiWNgGAWjYHACAyC2YGDgAbNtiNYiAdOSBhVMIFoLw2HCWszZD298XFEjkcffc/iY5I+C8/bm7GcPfub9YcPA396NVaNlT1qx4ZljEsUSZ9vSJCQMbifu7MlLluZJSGOQOHN2A1ZXHcgxk2xgk0hsOM9jJmFgcDsBJMLMk3CYwUAiF7uW82/Mfzb8k0icD9KSYHDOHigC0vIft5YbOWaMjW0SiRvO9phJHDA4wLjhBtiWA3i0PCuWbOyTSNx45liyZYNBcuKGG++SJeekJfPg9Mv55I0fG77ZJM47k3zw5o8/dkCH5R788MbGTo6/vRerFmyAh4GJBxZPRGth/EGK+lEwCkbBKBjuAADNfGO8tVhpaAAAAABJRU5ErkJggg==","orcid":"","institution":"Vancouver Prostate Centre","correspondingAuthor":true,"prefix":"","firstName":"Nader","middleName":"Al","lastName":"Nakouzi","suffix":""}],"badges":[],"createdAt":"2025-03-21 23:23:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6280571/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6280571/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":80291891,"identity":"e483ce28-89bc-4d72-866f-3ca388cce7d2","added_by":"auto","created_at":"2025-04-10 08:07:31","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1085072,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Wongetal.Mainfigures1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/df0a2ee237248416185fc101.jpg"},{"id":80293244,"identity":"cea3d2ee-afd9-45ab-a810-da30cf0b083e","added_by":"auto","created_at":"2025-04-10 08:15:31","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1400040,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Wongetal.Mainfigures2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/67cddcdd482289c2c4b29d4b.jpg"},{"id":80291895,"identity":"a3bed445-e04c-4b01-bbd8-e58302b4c3b5","added_by":"auto","created_at":"2025-04-10 08:07:32","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1386945,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Wongetal.Mainfigures3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/45b5c6ab39434bf913edbb9b.jpg"},{"id":80293886,"identity":"87bd8e6b-d984-43da-bbb7-e5f1e0f7bc6c","added_by":"auto","created_at":"2025-04-10 08:23:32","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":3556713,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Wongetal.Mainfigures5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/5c6d6df832ac843c425201d4.jpg"},{"id":80293245,"identity":"a419d449-b1d6-4457-8184-ec507abafa57","added_by":"auto","created_at":"2025-04-10 08:15:32","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":2431111,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Wongetal.Mainfigures7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/cc2f0cdb243325f6a21867c7.jpg"},{"id":80293246,"identity":"6639652d-4513-4b3e-b091-d1d0ef880ac9","added_by":"auto","created_at":"2025-04-10 08:15:32","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":2551331,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Wongetal.Mainfigures9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/8ad1627920086e87f832f7c4.jpg"},{"id":83288440,"identity":"86413ec7-bbfd-4dfe-a275-cd5e5a1cce19","added_by":"auto","created_at":"2025-05-22 12:17:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":13545699,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/85882bfa-7e8c-404e-bc06-43e369bf97d7.pdf"},{"id":80291896,"identity":"5cba211b-f008-4da3-8e79-883d744f47f5","added_by":"auto","created_at":"2025-04-10 08:07:32","extension":"xlsx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":536411,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalMaterialTable1and2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/25457557009b570ff4a04eb8.xlsx"},{"id":80293248,"identity":"db14770b-868a-489f-ad4d-21745200b467","added_by":"auto","created_at":"2025-04-10 08:15:32","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":588949,"visible":true,"origin":"","legend":"","description":"","filename":"Wongetal.Suppfigures.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6280571/v1/ecd1f9ab82c1a7f18ed02e05.pdf"}],"financialInterests":"Competing interest reported. N. AN., M.D., and N. F. are co-founders of SnapCyte Solutions Inc. S. A. and JM. S. are employed by SnapCyte Solutions Inc., and C. D., D. G., and JM. P. hold shares in SnapCyte Solutions Inc. The authors declare that these affiliations do not influence the objectivity or integrity of the research presented in this manuscript. All other authors declare no competing interests.","formattedTitle":"Enhanced Precision in Cell Culture Analytics: Leveraging Artificial Intelligence for Unbiased and Non-Destructive Assessment of Cell Growth and Viability.","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eRoutine quantification of confluency, cell counts, and cell viability are critical steps in life science research and biopharmaceutical industry workflows [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Accurate and consistent measurements of these parameters are essential for data reproducibility and experimental success. Confluency refers to the percentage of a surface area covered by a layer of cells in a culture vessel [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Regular determination of confluency is commonly used as a readout for cell growth and fitness in research fields such as cancer biology, stem cell research, and regenerative medicine [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In any experimental setup, maintaining optimal cell density is crucial, as high confluency can affect growth dynamics, delay cell passaging and harvesting, and impact downstream experiments. Overcrowding can limit nutrient access, reduce cell viability, and cause cell detachment from the culture plate surface [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Conversely, low confluency can result in insufficient cell-to-cell contact, impairing cell signaling and growth [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Growth dynamics and cell concentrations depend on inoculum cell count [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Cell counting is used for determining seeding density but also for tracking cell growth rates particularly in non-adherent cell cultures, providing information on cellular fitness and/or responses to experimental treatments. Obtaining precise cell numbers is required for seeding cells in culture plates to be used in subsequent experimental setups, normalizing results, and optimizing experimental conditions for procedures such as cell transfections [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. In clinical settings, precise cell counts are crucial for determining healthy ranges of cell populations, assessing toxicity, or evaluating immune reactions and other clinical parameters. Accurate cell counts are therefore vital for both life science research and clinical outputs [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTracking cell growth is instrumental for evaluating fitness of cell cultures and the proliferation of live cells serves as an important indicator of culture conditions. Changes in growth rates can provide information on cellular responses to stimuli or drug treatments. Given its importance, there are various methods for measuring cell growth. They can roughly be categorized into manual, colorimetric, and image-based methods. Examples of these methods include chemical dyes (\u003cem\u003ee.g.\u003c/em\u003e, MTT (3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide) assay), manual enumeration [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], automated cell counters [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], manual image analyses (\u003cem\u003ee.g.\u003c/em\u003e, Image J) [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], and automated image processing methods (\u003cem\u003ee.g.\u003c/em\u003e, Incucyte\u0026reg;) [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. While these techniques are valuable, the readouts are affected by human bias and variability in experimental conditions. Conventional determination of cell counts and cytometric parameters, most prominently, alive and intact dead cell ratios are performed using a hemocytometer (\u003cem\u003ee.g.\u003c/em\u003e, B\u0026uuml;rker-T\u0026uuml;rk counting chamber) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The major advantage of this conventional approach is the direct observation of the cell culture by the operator, which enables rapid detection of problems such as contamination or aggregation of cells. However, this method is prone to human error that can occur during many stages of the cell enumeration process, including mixing, handling, and dilution of the cells. Biggs et al. performed a study to quantify the standard error among trained technicians who counted red blood cells from one sample of blood by hemocytometers [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. They reported that the standard errors between the results of individual experiments conducted by two technicians and five technicians were 3.6% and 7.6% respectively, suggesting that the magnitude of errors increases with the number of individuals in an experiment. Different researchers might interpret mammalian cell boundaries in a hemocytometer differently, impacting results. A more recent study conducted by Manzini et al. showed that the variation can reach nearly 20% among different operators who are highly experienced in cell counting [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Another report documented the effects of dilution factors on total cell concentration measurements from manual methods. The coefficients of variation surged from 0.072 to 0.2 between 0.3 and 0.5 dilution fraction in an experiment, indicating the degree of human error in diluting samples during the cell counting process [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Additionally, the reproducibility of manual cell counts can be low, especially if high cell density cultures are used [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Beyond the issue of human error, manual counting of cells is also a time-consuming task, especially when having to perform multiple counts on the same culture for increased accuracy. This can lead to process fatigue and cause errors due to subjective interpretation of borderline cases.\u003c/p\u003e \u003cp\u003eColorimetric assays, such as MTT (3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide) and crystal violet, are widely used for assessing cell proliferation due to their extensive validation in numerous publications. However, these methods have notable limitations. The MTT assay measures metabolic activity by relying on the reduction of MTT by mitochondrial dehydrogenases to form an insoluble formazan product. This method is influenced by factors such as mitochondrial activity, which may not directly correlate with cell number or viability, leading to inaccuracies. Also, compromised metabolic activity in viable cells can yield misleading results [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Crystal violet staining, which involves binding of dye to DNA in viable cells, is semi-quantitative and influenced by cell size and DNA content variations. It does not distinguish between live and dead cells, potentially overestimating proliferation [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Both MTT and crystal violet assays are destructive, requiring cell lysis that prevents further analysis of the same cell population. Additionally, the definitive endpoint of the assay precludes real-time monitoring of growth and viability, calling for multiple parallel setups that increase the risk of error. The use of plate readers adds to the complexity and cost. Other reagent-based assays, such as WST-1 and CyQuant, offer some advantages but not without notable limitations. The WST-1 assay measures the conversion of a tetrazolium salt to a water-soluble formazan product, which is a more stable reaction compared to MTT. However, the assay still depends on cellular metabolic activity, which can vary independently of cell number [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. CyQuant assays use a fluorescent dye that binds nucleic acids providing a more direct and sensitive measurement of cell number. However, CyQuant requires cell lysis, limiting subsequent live-cell analyses [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Collectively, reagent-based methods for assessing cytostasis or cytotoxicity are often static or indirect, evaluating treatment effects at single time points or on associated cellular processes, such as membrane integrity. These methods cannot capture dynamic changes over time and they rely on endpoint measurements that do not reflect temporal variations in cell growth or viability. These limitations highlight the need for more accurate, dynamic, and non-destructive methods for assessing cell growth and viability in cell cultures.\u003c/p\u003e \u003cp\u003eImage-based methods measuring cell confluency (instrument-linked methods such as IncuCyte\u0026reg;) have gained popularity in recent years. While offering the convenience of automation and real-time tracking, these instruments are costly, requiring consumables and maintenance, and have limited user capabilities, making them inaccessible for many laboratories. Other image-based methodologies have also been developed using image processing, or more recently, machine learning (ML) to overcome user variability and reduce processing time, but they come with their own limitations. For example, the method developed by Soleimani et al. involves a multi-step process with image normalization, contrast enhancement, denoising, binary image conversion, modification, and finally measurement [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. This method's empirical modification step, akin to morphological opening [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], may lead to incorrect cell and noise classification, as manual adjustments in image processing can introduce bias and misclassification. Other methods struggle with extreme density scenarios. Wang et al. reported low accuracy (F-score\u0026thinsp;\u0026lt;\u0026thinsp;80%) for low confluency images (\u0026lt;\u0026thinsp;40%) and the method relied on artificial rules, such as thresholds, which hindered algorithm generalizability [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Furthermore, the deep learning-based approach presented by Ayanzadeh et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] shows great potential but focuses solely on low confluency cases, neglecting high confluency scenarios, thereby highlighting the need for improved methods. Overall, while current image-based methods offer advancements over reagent-based techniques, they require further development to address limitations and improve accuracy and accessibility for a wider range of laboratories. Transitioning from these traditional methods, deep learning-based computer vision emerges as a superior approach, setting new benchmarks in image-based cell analytics. Although there have been attempts to integrate various ML technologies into cell analysis, these methods require further enhancement to accommodate precise detection of clumped or overlapping cells, and adaptability to different cell morphologies and imaging conditions [\u003cspan additionalcitationids=\"CR31\" citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. For instance, the methodology developed by Jiang et al. employs a random forest and density map framework utilizing handcrafted features that limit its performance due to a limited receptive field [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Similarly, the approach by Schmidt et al. uses a density-regression deep convolutional neural network, combining a U-Net-like density predictor with a VGG-like regression model [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Although innovative, this approach struggles with detecting clumped and overlapping cells. Cellpose offers a robust segmentation algorithm that performs well across various cell types with minimal adjustments [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. However, this method falters under varying imaging conditions and morphological changes in our experiments, affecting its accuracy and limiting its general applicability to different data types. These limitations underscore the need for more robust and objective methods for cell growth quantification. Here, we describe the development and performance of a novel AI platform technology for advanced cell analytics that can be easily integrated with most laboratory workflows.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDeep Learning Optimization for Cell Confluency Detection\u003c/h2\u003e \u003cp\u003eCell growth patterns are reflected in numbers and densities of cells in a defined area or volume. These patterns can be captured in microscope images that are also particularly well suited for AI assessment. To build AI capacity able to assess growth patterns of cells, we designed a simple workflow that utilized an iterative human-in-the-loop approach for\u003c/p\u003e \u003cp\u003efine-tuning the confluency.Starting with a naive U-Net model, multiple training\u003c/p\u003e \u003cp\u003ecycles using human annotated and corrected sets were needed to ultimately reach output performances of \u0026gt;\u0026thinsp;90% accuracy (Fig.\u0026nbsp;1A). We first determined an AI model that could adequately report on cell density or confluency as a readout for cell growth. We employed a convolution-based encoder-decoder architecture based on U-Net for training of the AI \u003cb\u003e(Fig.\u0026nbsp;1B)\u003c/b\u003e. The U-Net architecture, characterized by its \"U\" shape, includes a contracting path, a bottleneck, and an expansive path, totaling 23 convolutional layers. Each segment of the architecture is critical for capturing features at different resolutions and effectively combining them for precise pixel-wise classification [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. We utilized Accuracy, Precision, Recall and F1-score parameters to assess the pixel-level segmentation performance of our model quantitatively. The Accuracy parameter provides a comprehensive measure of segmentation performance on both foreground and background [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e], while the F1-score focuses on the accuracy of target segmentation by the network [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Following four training iteration cycles, the model achieved a desirable threshold of Accuracy (0.94), and F1 (0.91) \u003cb\u003e(Fig.\u0026nbsp;1C)\u003c/b\u003e.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eRobustness of Confluency Detection Across Heterogeneous Imaging Conditions and Cell Types\u003c/h3\u003e\n\u003cp\u003eSince the SnapCyte\u0026trade; ML model is designed to be user-independent, with no need for users to set parameters, its universality is crucial. This means the model must perform effectively with images acquired in diverse settings and across a wide range of cell types. In the first step, we compiled and annotated a test set of 86 images, covering a broad spectrum of image qualities and sizes. This test set consisted of 43 published images of ATCC\u0026rsquo;s top 23 cell lines, available at both low and high confluency, and 43 images from 42 cell lines, as shown in 18 publications [\u003cspan additionalcitationids=\"CR38 CR39 CR40 CR41 CR42 CR43 CR44 CR45 CR46 CR47 CR48 CR49 CR50 CR51 CR52\" citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e]. SnapCyte\u0026trade; analyzed all images, generating masks that were compared to HQP annotation masks (\u003cb\u003eFig.\u0026nbsp;2A\u003c/b\u003e and \u003cb\u003eTables S1, S2\u003c/b\u003e). SnapCyte\u0026trade; achieved a strong correlation with manual measurements (\u0026#119877;2\u0026thinsp;=\u0026thinsp;0.98) and a slope of 0.96, reflecting near-perfect alignment. We further tested SnapCyte\u0026trade; against existing ML models using the ATCC image dataset (\u0026#119899;= 41), including a modified Cellpose model [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], the approach described in Lobsenz et.al [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e], and a widely accessible community-contributed method \u0026ldquo;Otsu-Based method\u0026rdquo; (ht\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003etps://github.com/insilicomab/cell_confluency)\u003c/span\u003e\u003cspan address=\"http://tps://github.com/insilicomab/cell_confluency)\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Our analyses demonstrated a strong correlation between manually assessed confluency values and those generated by SnapCyte\u0026trade; at both high and low cell density conditions, with an R\u0026sup2; of 0.98. In contrast, the Cellpose, Lobsenz, and GitHub models showed weaker correlations, with R\u0026sup2; values of 0.59, 0.42, and 0.43, respectively \u003cb\u003e(Fig.\u0026nbsp;2B)\u003c/b\u003e. We also assessed the error metrics where the Mean Absolute Error (MAE) of SnapCyte\u0026trade; was significantly lower at 2.8, indicating exceptional accuracy with minimal deviation from actual values. Conversely, the other models displayed considerably higher MAEs of 15.4, 20.5, and 36, suggesting less precise predictive capabilities and significant average errors. The Root Mean Squared Error (RMSE) for SnapCyte\u0026trade; was also low at 3.7, highlighting the model's accuracy and error consistency. The other models, such as GitHub, Lobsenz, and Cellpose, recorded RMSEs of 24.9, 28.4, and 40.8 respectively, indicating moderate to significant errors that could compromise model reliability and effectiveness in practical applications \u003cb\u003e(Fig.\u0026nbsp;2C)\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eFurthermore, we explored the model's robustness concerning image resolution by analyzing 10 images of diverse cell lines (e.g., MCF7, MG63, LNCaP, VCaP, IGRCaP-1, NCIH660) at 80%, 50%, 20%, and 8% of their original resolutions. The confluency value deviations between these resolutios were less than 2% \u003cb\u003e(Fig.\u0026nbsp;2D)\u003c/b\u003e, showcasing the model's robustness against variations in image quality. Overall, our data demonstrate that SnapCyte\u0026trade; maintains high robustness, universality, and independence from experimental settings such as microscope lighting or image resolution.\u003c/p\u003e \u003cp\u003e \u003cb\u003eOptimizing Image Sampling for Efficient and Accurate Confluency Estimation in Cell Culture Vessels at Various Magnifications\u003c/b\u003e \u003c/p\u003e \u003cp\u003eSince SnapCyte\u0026trade; relies on sample images to predict total vessel density, determining the minimal number of images needed for accurate measurements is essential. We investigated the time efficiency of capturing images across various cell culture vessels at different magnification levels to accurately estimate vessel confluency. At 4,000\u0026times; and 10,000\u0026times; magnification, the minimum number of images required to achieve a standard deviation of less than 5% for 96-well plates, 12-well plates, 6-well plates, and 10 cm dishes was found to be 2, 2, 4, and 5, respectively \u003cb\u003e(Fig.\u0026nbsp;3A, 3B)\u003c/b\u003e. This corresponds to 36.5%, 3.0%, 0.6%, and 0.3% of the total vessel area at 10,000\u0026times; magnification, and 22.8%, 18.9%, 13.8%, and 2.1% at 4,000\u0026times; magnification, respectively.\u003c/p\u003e \u003cp\u003eThe total imaging times for these vessels (all wells in a vessel) were 32, 4, 4, and 1 minute, respectively. However, for 96-well plates at 4,000\u0026times; magnification, a single image covering 11.4% of the well was sufficient to accurately predict the confluency of the entire well \u003cb\u003e(Fig.\u0026nbsp;3B\u003c/b\u003e, \u003cem\u003eupper panel\u003c/em\u003e), reducing the total imaging time to approximately 16 minutes for the whole vessel. Although images were acquired manually, the low number of required images per vessel and the rapid analysis by the ML model make SnapCyte\u0026trade; an efficient method for assessing confluency or density of cells in most basic laboratory settings. This efficiency could potentially be further enhanced by integration with an automated imaging system, offering even faster and more consistent results.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBenchmarking SnapCyte\u0026trade; Confluency Detection Against Standard Proliferation Assays\u003c/b\u003e Accurate and efficient monitoring of cell growth is essential for biological research and drug development studies. Traditional cell proliferation assays either track the number of cells over time or rely on surrogate readouts, such as metabolic activity, ATP or the percentage of area covered by adherent cells. We aimed to investigate the correlation between confluency measurements by SnapCyte\u0026trade; and cell counts. We first plated increasing concentrations of various cell lines, including epithelial (MCF7, PC3), mesenchymal (MG63), and fibrosarcoma (HT1080) cells, and assessed cell confluency 6 hours after plating using SnapCyte\u0026trade;. The correlation between cell counts and confluency was evaluated across three independent experiments. Results demonstrated a strong correlation between confluency assessed by SnapCyte\u0026trade; and cell count, with a coefficient of determination (R\u0026sup2;) greater than 0.9412 \u003cb\u003e(Fig.\u0026nbsp;4A; Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003eA, C, E, G).\u003c/b\u003e Next, we performed a similar analysis 24 hours after plating where wells seeded with 1.00E\u0026thinsp;+\u0026thinsp;06 cells had reached near-complete confluency (~\u0026thinsp;100%). The data showed a high linear correlation \u003cb\u003e(Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003eB, D, F, H).\u003c/b\u003e Linear regression analyses across all experiments, including four cell lines with three replicates at each time point, confirmed a robust correlation (average R\u0026sup2; = 0.974) \u003cb\u003e(Fig.\u0026nbsp;4B).\u003c/b\u003e Additionally, analysis of linear regression equations from replicates in each experiment revealed consistent slopes and interception points, underscoring the high reproducibility of the correlation (\u003cb\u003eFig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003eA-H\u003c/b\u003e). Collectively, these findings demonstrate that SnapCyte\u0026trade; can use cell confluency to precisely report on cell numbers and proliferation rates in 2D cell cultures across diverse cell lines.\u003c/p\u003e \u003cp\u003eTo further establish the performance and reliability of SnapCyte\u0026trade; as a tool for precise monitoring of cell proliferation, we tested its accuracy and correlation in measuring cell growth across various techniques and assessed its performance in terms of ease of use, time efficiency, and reproducibility. We benchmarked SnapCyte\u0026trade; against established reagent-based proliferation assays, including Crystal Violet (CV) staining, CyQuant, and WST-1. Initially, we evaluated the correlation between the normalized viability readouts from each assay and the initial cell counts at time of seeding. We found a high correlation across all techniques with no significant statistical difference in the slopes between SnapCyte\u0026trade; and the three alternative assays. Similarly, no significant difference was observed in SnapCyte\u0026trade; data collected by two independent users (F\u0026thinsp;=\u0026thinsp;0.53, dfn\u0026thinsp;=\u0026thinsp;4, dfd\u0026thinsp;=\u0026thinsp;35, P\u0026thinsp;=\u0026thinsp;0.71) (\u003cb\u003eFig.\u0026nbsp;4C\u003c/b\u003e). We also assessed the variability (delta) of normalized viability readouts among the different methods. The delta between SnapCyte\u0026trade; users was \u0026lt;\u0026thinsp;0.5%, as compared to 2% for the colorimetric CV assay and 4.5% for the metabolic assays (WST-1 and CyQuant) (\u003cb\u003eFig.\u0026nbsp;4D\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eIn a subsequent experiment, we employed SnapCyte\u0026trade; to monitor cell growth (measured as confluency over time) and compared these data against the conventional CV method (\u003cb\u003eFig.\u0026nbsp;4E\u003c/b\u003e). The data were similar with a delta value of 1.8% between the averages of the two methods (\u003cb\u003eFig.\u0026nbsp;4F\u003c/b\u003e). Additionally, no significant difference was observed between the two methods at any of the time points. To compare, SnapCyte\u0026trade; provided fast (15 min/time point), accurate, and time-lapsed growth data, facilitating direct use of the cell cultures in downstream applications. By contrast, the CV method necessitated multiple cultures for endpoint readouts and required more than 2 hours of processing time.\u003c/p\u003e \u003cp\u003eWe next evaluated the performance of SnapCyte\u0026trade; as a readout for cytotoxicity relative to the live cell imaging system IncuCyte\u0026reg; and the MTT assay. In this study, we assessed the cytotoxic effects of docetaxel on PC3 cells at two different concentrations. The experiments were conducted in triplicates and repeated by independent users. Results demonstrated that SnapCyte\u0026trade; generated comparable values to both IncuCyte\u0026reg; and the MTT assay, with SnapCyte\u0026trade; and IncuCyte\u0026reg; displaying lower standard deviations than the MTT assay \u003cb\u003e(Fig.\u0026nbsp;4G\u003c/b\u003e). Additionally, SnapCyte\u0026trade; exhibited less variability between independent experiments, further emphasizing its reliability and output reproducibility.\u003c/p\u003e \u003cp\u003eTo measure inter-user reproducibility, the same biological condition was plated four times in nine replicates and measured by four experienced users using IncuCyte\u0026reg;, SnapCyte\u0026trade;, and MTT. The data indicated no difference between techniques and users with SnapCyte\u0026trade; and IncuCyte\u0026reg; (standard deviation 14% and 13%, respectively), and showed that these methods achieved less variability as compared to the MTT assay (27%) (\u003cb\u003eFig.\u0026nbsp;4H\u003c/b\u003e). This underscores the advantages of non-invasive measurement methods in providing consistent, user-independent data. Moreover, when the same experiment was measured by four different users setting their own parameters, paired measurements showed an average standard deviation of 6.5% and 7.5% for SnapCyte\u0026trade; and IncuCyte\u0026reg;, respectively.\u003c/p\u003e \u003cp\u003eTo validate the performance of SnapCyte\u0026trade; in drug testing applications, we evaluated the half maximal inhibitory concentration (IC50) of docetaxel across three distinct cell lines: MCF-7, PC3, and IGR-CaP1. Each cell line was tested in three independent experimental setups, and the IC50 values were calculated using transformation and sigmoidal regression analyses. SnapCyte\u0026trade; demonstrated robust correlation coefficients for PC3 (0.935), MCF-7 (0.982), and IGR-CaP1 (0.972), with IC50 confidence intervals of [0.9\u0026ndash;1.5], [0.59\u0026ndash;0.75], and [1.5-2], respectively. These values were compared to those obtained with IncuCyte\u0026reg; and the MTT assay, which generated lower correlation coefficients and wider IC50 confidence intervals of 0.707 [0.7\u0026ndash;2.7], 0.919 [0.68\u0026ndash;1.13], 0.843 [1.4\u0026ndash;3.1] for IncuCyte\u0026reg; and 0.649 [0.7-3], 0.761 [0.24\u0026ndash;0.74], 0.865 [1.35\u0026ndash;2.85] for the MTT assay in the respective cell lines \u003cb\u003e(Fig.\u0026nbsp;4I, J).\u003c/b\u003e\u003c/p\u003e \u003cp\u003eLastly, we evaluated the SnapCyte\u0026trade; model's ability to replicate published data by determining the doubling time of three cell lines across three independent experiments. Our results not only aligned closely with each cell line\u0026rsquo;s published doubling time, but also showed that the doubling time values were near the median of values from independent studies (\u003cb\u003eFig.\u0026nbsp;4K, L\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eCombined, the results indicate that SnapCyte\u0026trade; provides highly accurate and reproducible measurements of cell growth and confluency across diverse cell lines. The system\u0026rsquo;s non-invasive and time-lapsed capabilities make it an efficient and reliable alternative to conventional instrument and reagent-based assays. SnapCyte\u0026trade; offers comparable accuracy, reduced variability, and faster processing times and is well-suited for routine growth and survival assessments in cell biology laboratories.\u003c/p\u003e\n\u003ch3\u003eOptimization of Deep Learning Models for Cell Counting and Viability Assessment\u003c/h3\u003e\n\u003cp\u003eAccurate cell counts are essential for cell biology assays, particularly for non-adherent cells where confluency measurements are inadequate. We next aimed to optimize an ML model for assessing cell numbers, employing cellpose model [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]a standard U-Net architecture pre-trained on a diverse range of cellular images \u003cb\u003e(Fig.\u0026nbsp;5A)\u003c/b\u003e. We used various cell lines, beads, PBMCs, and RBCs that were loaded onto a hemocytometer to acquire images for annotation. The ML model utilized a combination of down-sampling and up-sampling techniques to process feature maps.\u003c/p\u003e \u003cp\u003eTraining and fine-tuning the Cell Count model using our human-in-the-loop approach \u003cb\u003e(Fig.\u0026nbsp;1A)\u003c/b\u003e required seven iterative cycles to achieve Precision and Recall rates of over 95% in detecting cells while excluding debris \u003cb\u003e(Fig.\u0026nbsp;5B)\u003c/b\u003e. However, further improvements posed challenges due to a 5% variation in labeling, which may result from discrepancies even among cell annotation experts. A drop in the F1 score during cycle 4 was attributed to the introduction of a more complex and variable dataset, reflecting more realistic scenarios. After cycle 6, our analysis indicated that annotation inconsistencies were the primary obstacle for further improvements, as supported by previous research by Kang et al. [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e]. A dataset review identified and resolved these inconsistencies, leading to improved performance in the final cycle, achieving 95% Precision and Recall at cycle 7.\u003c/p\u003e \u003cp\u003eWe further evaluated the performance of SnapCyte\u0026trade; in segmenting and counting single cells against other state-of-the-art models (Cyto1 [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], Cyto3 [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e], Omnipose [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e], and StarDis [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] on a randomly selected dataset of 50 images. The results reveal that SnapCyte\u0026trade; consistently outperforms the alternatives across all three key metrics: precision, recall, and F1-score \u003cb\u003e(Fig.\u0026nbsp;5C)\u003c/b\u003e. SnapCyte\u0026trade; achieves the highest average precision of 94%, significantly surpassing Cyto1 and Cyto3 (both at approximately 75%), and far exceeding Omnipose (37%) and StarDist (48%). Similarly, the recall of SnapCyte\u0026trade; is 97%, considerably higher than Cyto1 (72%) and Cyto3 (78%), with Omnipose (63%) and StarDist (42%) demonstrating inferior performance. Finally, SnapCyte\u0026trade; achieves the highest F1-score at 95%, reflecting its superior balance between precision and recall, compared to Cyto1 (73%), Cyto3 (74%), Omnipose (49%), and StarDist (45%). Among the tested models, Cyto1 and Cyto3 demonstrate moderate and consistent performance across all metrics, though they lag behind SnapCyte\u0026trade;. Omnipose exhibits substantial deficiencies in precision and F1-score, likely due to its reduced ability to accurately detect individual cells. StarDist shows poor overall performance in this diverse data set, with significantly low precision and recall, indicating challenges in both cell detection and segmentation. These results highlight the robustness and accuracy of SnapCyte\u0026trade;, establishing SnapCyte\u0026trade; as the most reliable tool for cell counting among the models tested.\u003c/p\u003e \u003cp\u003eTo assess model performance across various cell densities, we evaluated Precision, Recall, and F1 scores using subsets of our dataset compiled based on annotated cell numbers per image. We analyzed 12 images with fewer than 50 cells, 12 images with 50 to 300 cells, and 12 images with more than 300 cells. The data showed consistent and strong performance across all groups \u003cb\u003e(Fig.\u0026nbsp;5D)\u003c/b\u003e. As the number of cells increased, the metrics (Precision, Recall, and F1) improved, reaching 0.95 for images with more than 50 cells. Although we did not determine an upper limit for cell numbers, the model\u0026rsquo;s robustness in denser images suggests its suitability for applications involving cell clumping and overlap.\u003c/p\u003e \u003cp\u003eGiven the importance of precise cell size measurements in biological assays, particularly for distinguishing cell types, we optimized a Size Estimation model, which works in conjunction with the Cell Count model. The process involves two steps: first, the Cell Count model evaluates images using a default diameter to compute a Style array\u0026mdash;256 float values array, representing image features. A linear regression model then predicts cell size based on the aforementioned Style array. Finally, the Cell Count model generates output-masks using the median of the cell sizes, as determined by the Size Estimation model. The image is resized according to the predicted diameter, and the Cell Count model generates output masks, from which the median object size is determined as the final predicted size. Utilizing a visual representation where the x-axis represents the true cell diameter and the y-axis the predicted cell diameter (in pixels), we observed a close alignment between predicted and actual sizes (R\u0026sup2; = 0.95) (\u003cb\u003eFig.\u0026nbsp;5E\u003c/b\u003e). This strong correlation confirms the accuracy of the Size Estimation model, making it a reliable tool for downstream applications requiring precise cell size measurements. The combination of robust predictions from the Cell Count model and linear regression ensures accurate size estimations with low error rates suitable for most cell counting requirements.\u003c/p\u003e \u003cp\u003eIn many cell biology assays, cell count is meaningful only when combined with viability assessments. Thus, we aimed to develop an ML model capable of distinguishing live and dead cells based on Trypan Blue or Erythrosin-B staining, two commonly used dyes for cell viability assessment [\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e]. To determine cell viability, images of cells stained with these dyes first undergo analysis by the Cell Count model to identify individual cells. The resulting segmentation mask, combined with the original image, is then processed through a separate U-Net network. This U-Net architecture comprises four convolutional layers with ReLU activations and max pooling for feature extraction, followed by three transposed convolutional layers for spatial restoration. The final layer employs a convolutional operation with a Tanh activation function to generate a single-channel segmentation mask for accurate viability detection. The model was trained on a dataset of 244 images, with an additional 244 augmented images to enhance robustness. Our validation and test set include 62 and 77 original images respectively. Evaluation metrics, including Accuracy, Precision, Recall, and F1-score, are depicted in \u003cb\u003e(Fig.\u0026nbsp;5F\u003c/b\u003e), alongside an example output image demonstrating the accuracy of the viability detection.\u003c/p\u003e \u003cp\u003eAmong the available machine learning models for cell counting and segmentation, to our knowledge, only the model published by Kuijpers et al [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] was reported to distinguish live and dead cells using trypan blue staining. Using a dataset of 20 images of trypan blue-stained cells loaded into different types of chambers (e.g., hemocytometers and KOVA slides), we compared SnapCyte\u0026trade; to the Kuijpers model for total cell count and live/dead cell discrimination. For total cell count, Kuijpers exhibited high error rates (MAE: 57.4%, MSE: 3656, RMSE: 60.5%), indicating consistently large errors, with occasional extreme outliers contributing to the higher RMSE. In contrast, SnapCyte\u0026trade; achieved significantly lower errors (MAE: 2.4%, MSE: 26.7, RMSE: 5.2%), demonstrating both accuracy and consistency. Similarly, for live and dead cell classification, SnapCyte\u0026trade; maintained minimal errors (MAE: 2.3%-2.6%, RMSE: 3.1%-4.2%), compared to Kuijpers' higher errors (MAE: 34.2%-23.8%, RMSE: 42.6%-28.4%) (\u003cb\u003eFig.\u0026nbsp;5G\u003c/b\u003e). Although the Kuijpers model had false positive and false negative rates below 2% for live/dead classification, its higher overall errors stemmed from poor cell segmentation, highlighting the importance of accurate segmentation. The low and closely aligned MAE and RMSE across all tasks reflect precise and reliable predictions and confirm robustness and superior performance of SnapCyte\u0026trade; in cell analyses.\u003c/p\u003e \u003cp\u003eAccurate detection and counting of particles of different sizes is critical when working with mixed populations of cells or particles, such as those commonly used in co-cultures, primary cultures, and patient samples. To ensure that our model could handle diverse sample types, we tested its capability to count shapes of different sizes. We used 6\u0026micro;m, 10\u0026micro;m, and 16\u0026micro;m flow cytometer size reference beads, which were diluted to different concentrations or mixed in a 1:1:1 ratio. The number of beads in each sample was compared between SnapCyte\u0026trade; and manual counting. Our results showed a strong correlation between the SnapCyte\u0026trade; values and expected concentrations for all three bead sizes, both separately and when mixed together (\u003cb\u003eFig.\u0026nbsp;5H-K\u003c/b\u003e). Specifically, the coefficient of determination for the 16\u0026micro;m beads was as high as 0.9999 and \u0026gt;\u0026thinsp;0.99 for 6\u0026micro;m and 10\u0026micro;m beads. This indicates that our model can accurately detect and count homogeneous and heterogeneous samples with different particle sizes, highlighting its robustness in handling mixed cell populations (\u003cb\u003eFig.\u0026nbsp;5K\u003c/b\u003e).\u003c/p\u003e\n\u003ch3\u003eComparative Performance of the SnapCyte™ Cell Count Model and Conventional Cell Counting Methods\u003c/h3\u003e\n\u003cp\u003eWe first tested the SnapCyte\u0026trade; size estimation model using a diverse set of biologically relevant samples. A dataset of 96 images of human red blood cells (RBCs), human peripheral blood mononuclear cells (PBMCs), PC3, and MCF7 cells, was employed. The predicted values of SnapCyte\u0026trade; were comparable to absolute cell sizes determined by scientists \u003cb\u003e(Fig.\u0026nbsp;6A)\u003c/b\u003e. The results demonstrated a strong correlation between SnapCyte\u0026trade; and the actual cell sizes for all three types of samples (R\u0026sup2; = 0.9733) \u003cb\u003e(Fig.\u0026nbsp;6B)\u003c/b\u003e. Furthermore, the Accuracy, Recall, and F1-score of the model were consistently above 90%, regardless of the cell type analyzed \u003cb\u003e(Fig.\u0026nbsp;6C)\u003c/b\u003e. These findings highlight the ability of SnapCyte\u0026trade; to accurately determine and differentiate cell sizes across various sample types, including heterogeneous PBMC samples.\u003c/p\u003e \u003cp\u003eNext, we benchmarked the SnapCyte\u0026trade; model\u0026rsquo;s counting and viability capabilities against manual assessments performed by experienced scientists. We used a test dataset of 128 images featuring Trypan Blue- or Erythrosin B-stained PC3 and MCF7 cells with varying viability percentages. Cells were loaded onto hemocytometers or KOVA slides for evaluation by the SnapCyte\u0026trade; viability module. Results showed that the average absolute difference between SnapCyte\u0026trade; and manual counting was less than 5% for both Trypan Blue and Erythrosin B staining across different live/dead cell ratios \u003cb\u003e(Fig.\u0026nbsp;6D, 6E\u003c/b\u003e). Importantly, the model\u0026rsquo;s performance was independent of the slide type when tested on KOVA slides (\u003cb\u003eFig. \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e\u003c/b\u003e). This demonstrates that the SnapCyte\u0026trade; viability model is a reliable tool for quantitative microscopy, providing accurate viability measurements for cells stained with two commonly used viability dyes.\u003c/p\u003e \u003cp\u003eMost automated cell counters have a reliable detection range of 1.00E\u0026thinsp;+\u0026thinsp;05 to ~\u0026thinsp;1.00E\u0026thinsp;+\u0026thinsp;07 cells/mL [\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e]. This range limitation sometimes requires cells to be concentrated or diluted to obtain accurate counts. We tested the SnapCyte\u0026trade; model\u0026rsquo;s performance across a wide range of cell concentrations to establish the dynamic range. Data showed that SnapCyte\u0026trade; provided reliable cell counting in the range of 1.00E\u0026thinsp;+\u0026thinsp;04 cells/mL to 2.50E\u0026thinsp;+\u0026thinsp;07 cells/mL \u003cb\u003e(Fig.\u0026nbsp;6F\u003c/b\u003e), as indicated by a high coefficient of determination (R\u0026sup2; = 0.9847), comparable to Bio-Rad (R\u0026sup2; = 0.9953) and manual counting (R\u0026sup2; = 0.9978). We further assessed the ability of SnapCyte\u0026trade; to count cells in serially diluted PC3 cell samples. High-resolution images revealed that each cell was effectively segmented even in the highest concentration of 2.50E\u0026thinsp;+\u0026thinsp;07 cells/mL (\u003cb\u003eFig.\u0026nbsp;6G\u003c/b\u003e). This demonstrates that SnapCyte\u0026trade; accurately distinguishes individual cells in densely packed samples that are difficult to count in manual settings.\u003c/p\u003e \u003cp\u003eFinally, we evaluated the reproducibility of SnapCyte\u0026trade; compared to manual counting. Four researchers manually counted MCF7 cells at three concentrations (5E\u0026thinsp;+\u0026thinsp;06 cells/mL, 1.6E\u0026thinsp;+\u0026thinsp;06 cells/mL, and 2.9E\u0026thinsp;+\u0026thinsp;06 cells/mL) using a hemocytometer. In parallel, each researcher determined the cell count using SnapCyte\u0026trade;. As a reference, cell counts were also measured using the TC-20 cell counter. SnapCyte\u0026trade; results showed no statistically significant differences between SnapCyte\u0026trade; and manual counting measurements. It also showed smaller standard error deviations for SnapCyte\u0026trade; measurements compared to manual counting \u003cb\u003e(Fig.\u0026nbsp;6H\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eAll together our data demonstrate that SnapCyte\u0026trade; provides a robust, accurate, and efficient ML-based solution for assessing cell confluency, counting, and viability across diverse cell types and experimental conditions. Its U-Net architecture achieves\u0026thinsp;\u0026gt;\u0026thinsp;90% accuracy for confluency detection, with strong reproducibility and minimal deviations under varying image qualities and densities. SnapCyte\u0026trade; also demonstrated high correlation with traditional proliferation assays and outperformed manual counting methods, extending its dynamic range to 2.50E\u0026thinsp;+\u0026thinsp;07 cells/mL. The system\u0026rsquo;s non-invasive, time-efficient, and highly reproducible measurements, make it a reliable alternative to conventional assays, offering significant advantages for routine cell biology workflows and drug development studies.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eScientists face numerous challenges related to basic cell analytics proficiency. Research costs have increased drastically, the demand for faster methods have intensified, and the scientific community has encountered a reproducibility and replicability crisis [\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e], which reflects a multifactorial phenomenon where standardization and accuracy of technologies play significant roles.\u003c/p\u003e \u003cp\u003eIn cell analytics, traditional methods, such as manual cell counting and reagent-based approaches, are time-consuming, prone to human error, and require costly instrumentation. Additionally, recent image processing technologies have demonstrated limitations, including being time-consuming and requiring users to define subjective parameters (\u003cem\u003ee.g.\u003c/em\u003e, ImageJ) or they are inaccessible to many research groups due to cost. There is a critical need in cell biology research to eliminate user variability inherent in conventional methods and to develop accessible and reliable solutions to ensure data comparability between studies. AI-based systems have emerged as transformative tools in cell analytics, promising rapid, unbiased, and consistent cell analysis readouts. Recent advancements in automation and ML for basic cell culture tasks have shown promise, but only a few have proven to be accurate and applicable to a wide range of cells and experimental conditions. Many ML models currently available are trained on open-source datasets that feature idealized, clean images, often lacking the variability encountered in everyday cell culture environments. These datasets typically do not reflect the real-world challenges, such as debris, clumps, non-uniform lighting, or irregularities in sample preparation, making their models less effective in practical applications. Consequently, there is still an imminent demand for accurate and affordable deep learning models to be widely deployed in research.\u003c/p\u003e \u003cp\u003eSnapCyte\u0026trade; was developed using a rigorously curated dataset encompassing a wide array of cell lines and conditions, meticulously annotated to customize AI models for detection and analysis of cells across diverse experimental setups. We have customized three ML models to measure cell confluency, cell count, and cell viability\u0026mdash;three readouts that constitute the basics of cell analytics in a cell biology or life science laboratory. SnapCyte\u0026trade; achieves high precision and recall for all three models across all cell lines and image qualities that were tested, reflected by high F1 scores. Notably, for the Cell Count ML model, accuracy increased in samples containing a higher number of cells, proving the superiority of this method in counting concentrated samples. This is important, as dilution affects the prediction accuracy of total cell counts [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSnapCyte\u0026trade; consistently outperforms existing publicly available models with exceptionally low error metrics, demonstrating high precision and reliability. While some models achieve a balance of accuracy and consistency, others exhibit significant errors across all metrics, indicating a need for substantial adjustments or alternative modeling approaches to achieve acceptable predictive accuracy. This highlights the importance of customization in machine learning models. Generalist models like Cellpose are designed for broad applications, which can limit their precision, whereas SnapCyte\u0026trade; was developed using specialized and diverse datasets, ensuring superior accuracy and performance.\u003c/p\u003e \u003cp\u003eConfluency measurements are crucial for ensuring the consistency and reliability of cell culture experiments. Traditional methods often rely on subjective visual assessments, leading to significant errors and variability. Studies have shown that confluency eyeballing errors can be very high, especially at higher confluency levels, impacting experimental outcome and reproducibility. For example, a study published by Lin et al., discusses the challenges in accurately determining cell confluency, highlighting that visual estimations can greatly vary among researchers, especially among those with less experience. This variability can exceed 30%, affecting the consistency of experimental results [\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e]. High confluency levels can lead to contact inhibition, altering gene expression involved in cell cycle regulation and apoptosis, thus affecting studies on cell growth, migration, and response to treatments. Our validation tests of the confluency model on independent datasets, which included a high number of cell lines and random images from publications, demonstrated the universality of the ML model and its applicability to a wide range of laboratory settings (lighting, image resolution, and cell models). The F1 score remained above 91%, and the average deviation from manual masks was 3%.\u003c/p\u003e \u003cp\u003eThe SnapCyte\u0026trade; confluency model is highly relevant for data standardization in cell biology research. It enables scientists to instantly evaluate seeding homogeneity and report it accurately. This is crucial for experiments that are sensitive to seeding variation, such as genome editing and differentiation studies, thereby enhancing replicability and reproducibility. This model also assists scientists in maintaining healthy cell cultures by ensuring that cells are seeded homogeneously and passaged at adequate confluency to prevent genetic drift in cell cultures.\u003c/p\u003e \u003cp\u003eAdditionally, SnapCyte\u0026trade; allows for non-invasive, time-lapse cell growth measurement from the same vessel, enabling scientists to determine the optimal timing for treatments or cell editing. This capability helps avoid potential issues related to over or under-seeding, which can significantly affect experimental outcomes.\u003c/p\u003e \u003cp\u003eBalancing the positive considerations, it is important to acknowledge the limitations of image-based analysis, particularly when based on cell confluency. In some conditions, treatments or cell manipulations may lead to morphological changes and alterations in cell attachment, which can affect cell size and confound confluency measurements. This could result in cell growth assessment errors. Nevertheless, because the analysis is image-based, scientists can visually inspect cell morphology and detect changes, thereby mitigating the risk of misleading information.\u003c/p\u003e \u003cp\u003eAccurate cell counting is equally critical in cell biology research. Manual counting methods are prone to significant errors, while newer automated counters struggle with complex samples containing cells of different sizes. SnapCyte\u0026trade; addresses these challenges by accurately counting particles of various sizes, as demonstrated in correlation studies with manual counting (r\u0026thinsp;\u0026gt;\u0026thinsp;0.99). Notably, this model can count heterogeneous samples containing beads of varying sizes (6, 10, and 16 \u0026micro;m), further validated by its ability to distinguish different sizes accurately, including distinct populations of RBCs, PBMCs, and cancer cell lines. Additionally, SnapCyte\u0026trade; exhibits exceptional accuracy in detecting cells stained with Trypan Blue or Erythrosine B to assess viability. As many laboratories transition away from Trypan Blue staining due to its toxicity, the ability to use Erythrosine B provides a safer alternative. This versatility makes SnapCyte\u0026trade; an impactful tool for measuring cell viability in complex biological samples.\u003c/p\u003e \u003cp\u003eSince SnapCyte\u0026trade; uses images to analyze cells, an important consideration is how representative the images are of the entire vessel especially in the context of cell proliferation. Like many other techniques in cell culture data analysis, cell confluency is based on image sampling, and in the case of adherent cells, the homogeneity of the culture is a cornerstone for performing accurate and reproducible data analysis. Assuming homogeneous seeding, SnapCyte\u0026trade; requires a minimal sampling number (\u003cem\u003ei.e.\u003c/em\u003e 1\u0026ndash;2 for a 96-well plate, 2\u0026ndash;4 for 6- and 12-well plates and a 10 cm vessel). This ensures that SnapCyte\u0026trade; analyses are fast, and return data within a few minutes if using manual imaging. The ability of SnapCyte\u0026trade; to process data within minutes without necessitating invasive procedures aligns with the pressing need for versatile analytical tools that are compatible with downstream processes.\u003c/p\u003e \u003cp\u003eWe also investigated whether the cell confluency readout from SnapCyte\u0026trade; could be used as a surrogate readout for cell proliferation and cytotoxicity assays. First, the correlation between cell counts and SnapCyte\u0026trade; confluency was \u0026gt;\u0026thinsp;0.95 in all conditions tested, indicating that confluency could predict cell number and, hence, proliferation (\u003cb\u003eFig.\u0026nbsp;4A and Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e). When compared to reagent-based assays, results were comparable to standard methods with a low delta variation. When compared to Crystal Violet, the most commonly used manual assay for confluency, the variation was lower than 4% in all conditions. This offers the added advantage of avoiding errors related to Crystal Violet staining steps and reducing seeding-related variability between vessels by acquiring data from the same vessel. This approach also reduces the use of toxic reagents, and saves time by allowing immediate use of cells in downstream assays.\u003c/p\u003e \u003cp\u003eCompared to the IncuCyte\u0026reg;, data obtained on SnapCyte\u0026trade; were similarly accurate. Also, data between different SnapCyte\u0026trade; users were more consistent as parameter settings were predefined. Furthermore, when compared to MTT and IncuCyte\u0026reg;, IC50 values of cell lines treated with docetaxel and calculated with SnapCyte\u0026trade; were more accurate, as evidenced by a high non-sigmoidal correlation coefficient (\u003cb\u003eFig.\u0026nbsp;4I\u003c/b\u003e). As such, SnapCyte\u0026trade; produced similar IC50 values as other techniques but with smaller confidence intervals. While SnapCyte\u0026trade; is not tied to an automation system, manual image acquisition remained within acceptable time limits (15 minutes). SnapCyte\u0026trade; allows acquisition from any type of vessel and microscope, whereas many other methods and techniques are tied to their own instruments (e.g., IncuCyte\u0026reg;). Moreover, the phone app version of SnapCyte\u0026trade; transforms a smartphone into a cell analytics platform that can perform instant analyses on the device. Additionally, SnapCyte\u0026trade; reduces contamination risks by eliminating the need to transfer cells between different rooms for measurement, maintaining the integrity of cell cultures and experimental conditions. Finally, the last validation involved comparing SnapCyte\u0026trade;-generated data with published values of doubling time, demonstrating the accuracy of SnapCyte\u0026trade; by producing values close to the median and demonstrating highly reproducible results.\u003c/p\u003e \u003cp\u003eFor the cell count readout, SnapCyte\u0026trade; uses a hemocytometer but can also be used with any type of slide. It eliminates the need for additional machines and specific slides. SnapCyte\u0026trade; has a wide dynamic range, allowing quantification of highly concentrated samples and different sizes of cells. The data from SnapCyte\u0026trade; is user-independent, and the variation is minimal.\u003c/p\u003e \u003cp\u003eThe validation across a range of cell lines, culture conditions, and analytical parameters attests to the robustness and versatility of the SnapCyte\u0026trade; system. Future studies should focus on expanding the applicability of SnapCyte\u0026trade; to other complex biological systems and exploring its integration with other cutting-edge analytical tools. By offering high precision, efficiency, and user-independence, the SnapCyte\u0026trade; AI tool paves the way for more reproducible, accurate, and insightful scientific research, ultimately accelerating the pace of discovery and innovation in life science research.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eModel architecture\u003c/h2\u003e \u003cp\u003eIn this study, we evaluated multiple deep learning architectures to determine their suitability for cell detection and segmentation tasks. Models such as U-Net, Mask R-CNN, and YOLO were tested for their ability to identify cellular features in microscopy images. Initial evaluations revealed that U-Net and Mask R-CNN faced challenges in accurately detecting specific features in complex microscopy data. While YOLO showed efficiency in fast object detection, its precision was inadequate for the requirements of life science imaging [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWe tested several other models, including those utilized in literature and competition datasets, such as those from Kaggle, which often required specific image types (e.g., grayscale or single-cell fluorescent images). For instance, DeepCell [\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e] was designed for single-cell fluorescent microscopy images, while CellProfiler [\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e] underperformed on our multi-cell culture images, likely due to its instrument-specific limitations. DeepLab demonstrated promising results in general semantic segmentation tasks, but U-Net emerged as the most effective for cell microscopy, owing to its use of skip connections and its symmetrical structure, which improved feature extraction and boundary delineation [\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e]. This architecture proved especially beneficial for estimating confluency levels in complex cellular environments. Hence, we chose U-Net as the core architecture for our cell confluency model.\u003c/p\u003e \u003cp\u003eWe integrated and fine-tuned the Cellpose model [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], which was initially pretrained for general cell segmentation. While Cellpose demonstrated robustness and adaptability across various cell types, we identified the need for further fine-tuning to align with our specific application, enhancing precision in cell counting and segmentation under diverse imaging conditions.\u003c/p\u003e \u003cp\u003eWe performed a comprehensive evaluation of several regression models for cell size estimation, including Support Vector Regressor, Elastic Net Regressor, K-Nearest Neighbors Regressor, Random Forest Regressor, Gradient Boosting Regressor, and Linear Regressor. We conducted experiments using our dataset of images and compared the Mean Squared Error (MSE) across all models, yielding values of 43.5, 40.1, 36.1, 23.6, 12, and 10.9, respectively. Our analysis, based on these experiments, revealed that Linear Regression exhibited greater robustness when applied to unseen images from our dataset, outperforming the other models. Building upon Cellpose's Linear Regression Size Model, we integrated this Linear Regression model into our pipeline for cell size estimation. This addition ensures precise determination of cell diameter size, thereby enhancing the reliability and performance of our cell counting system across a wide range of imaging conditions.\u003c/p\u003e \u003cp\u003eIn our efforts to develop an optimal solution for detecting cell viability, we initially utilized unsupervised ML clustering algorithms to differentiate between live and dead cells in cell culture images. Unsupervised learning, beneficial in contexts without explicitly labeled data, allows algorithms to independently identify patterns or clusters within the data. However, these algorithms often classify images into two clusters, even when only live or dead cells are present. Among the unsupervised algorithms tested, K-means demonstrated acceptable performance with Trypan Blue-stained images but was less effective with other stains, such as Erythrosin B. Consequently, we decided to employ supervised learning techniques and chose the UNet architecture for viability detection. We conducted further experiments to optimize the UNet model's architecture and parameters, which resulted in satisfactory performance across various types of images.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eData Sets\u003c/h3\u003e\n\u003cp\u003eCell images were captured utilizing a microscope adaptor and a variety of cell phones (iOS and Android systems) and microscopes (Nikon ECLIPSE Ts2 inverted microscope, Leica DM1000 LED, Helmut Hund Wilovert AFL 30 Series 8 Inverted Trinocular Microscope) with 4\u0026times;, 10\u0026times;, and 20\u0026times; optical magnification, along with different digital magnifications. For the cell confluency dataset, 1500 images of multiple cell lines cultured in various vessels were captured and annotated. For the cell count dataset, various cell lines (PC3, MCF7, Raji,) and cell samples (RBCs, PBMCs) were cultured, adherent cells were detached, and cells were loaded into a hemocytometer or KOVA slides at different concentrations, both in the presence and absence of Trypan Blue or Erythrosin B. Cell death was induced by heat shock at 80℃ for 15 minutes. The dataset also included images of red blood cells, peripheral blood mononuclear cells (PBMCs) isolated as previously described [\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e], and beads of various sizes (6, 8, 10, and 16 \u0026micro;m) ( ThermoFisher; Cat# C16506 and Spherotech Inc; Cat# PPS-6K). Two thousand one hundred images were annotated by experienced scientists to generate cell count and live/dead masks for ML. To enhance the model's generalization capabilities, particularly for out-of-focus images, we implemented a Gaussian Blur technique on a subset of randomly chosen images.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eCell lines and cultures\u003c/h2\u003e \u003cp\u003eMCF-7, PC3, HELA, LNCaP, MG63, and HT1080 cells were procured from ATCC (Manassas, VA, USA). IGR-CaP1 cells were provided by Dr. Chauchereau (Gustave Roussy Institute, France) [\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e]. Cells were maintained in their appropriate media supplemented with 10% FBS in a humidified incubator at 37\u0026deg;C with 5% CO\u003csub\u003e2\u003c/sub\u003e. All cells were tested for mycoplasma regularly.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eIncuCyte\u0026reg; Analysis\u003c/h2\u003e \u003cp\u003eCell confluency was measured by IncuCyte\u0026reg; according to the manufacturer's protocol. Briefly, MCF7 cells, PC3 cells, and IGR-CaP1 cells were seeded in triplicate in 96-well plates. Cells were treated with Docetaxel at different concentrations and were then monitored on the IncuCyte\u0026reg; Live Cell Analysis System (Sartorius, USA). Images were taken at 4\u0026times; magnification and analyzed by IncuCyte\u0026reg; Live-Cell Analysis Systems.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eSnapCyte\u0026trade; Count\u003c/h2\u003e \u003cp\u003eNumber and viability of cells were determined by SnapCyte\u0026trade; Count and SnapCyte\u0026trade; Viability modules, respectively. Briefly, cells were cultured in various cell culture vessels and cell pellets were collected. 10 \u0026micro;L of each sample was individually loaded on a Hausser Scientific\u0026trade; Bright-Line\u0026trade; Phase Hemacytometer (Cat.# 026716) or KOVA\u0026trade; Glasstic\u0026trade; Slide (fisher scientific; Cat.# 22-270141) and the cell numbers were determined by SnapCyte\u0026trade; Count module. For cell viability analyses, equal volumes of Trypan Blue solution, 0.4% (Gibco; Cat.# 15250061) or 0.1% of Erythrosin B (ThermoFisher; Cat.#A14180.14) were added to the resuspended cells in corresponding cell culture media supplemented with 10% FBS before the samples were loaded on a hemocytometer or KOVA\u0026trade; Glasstic\u0026trade; Slide (fisher scientific; 22-270141). Cell viability was determined by SnapCyte\u0026trade;. For beads analysis, 10 \u0026micro;L of the Cell Sorting Set-up Beads (for UV lasers) (ThermoFisher; Cat.# C16506) or Polystyrene Particle Size Standard Kit, Flow Cytometry Grade (Spherotech Inc; Cat.# PPS-6K) were prepared at different concentrations with distilled water and loaded on a hemocytometer. Bead numbers were determined by SnapCyte\u0026trade; Count module. Four photos per sample were taken by an Apple iPhone 8 and an adapter provided by SnapCyte\u0026trade;, mounted on a Nikon ECLIPSE Ts2 inverted microscope at 10\u0026times; magnification. Region of Interest (ROI) was set up for each photo taken.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eSnapCyte\u0026trade; Confluency\u003c/h2\u003e \u003cp\u003eConfluency of cells was measured by SnapCyte\u0026trade; Confluency according to the manufacturer's protocol. Briefly, cells were cultured in various cell culture vessels and cell confluency was determined at different timepoints by SnapCyte\u0026trade; Confluency in photos taken with various phones and SnapCyte\u0026trade; universal Smartphone Adapter mounted on Nikon ECLIPSE Ts2 inverted microscope. Region of Interest (ROI) was set up for each photo taken and results were exported and analyzed. Microscope magnification used and the numbers of photos taken at different timepoints in experiments are described in corresponding figure legends.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eColorimetric assays\u003c/h2\u003e \u003cp\u003eColorimetric assays (\u003cem\u003ei.e.\u003c/em\u003e MTT, WST1, Cyquant, and Crystal violet assays) were used to determine the viability of MCF-7, PC3 and/or IGR-Cap1 at 72 hours after treatment of Docetaxel, and subsequently half maximal inhibitory concentration (IC50). Measurements by colorimetric assays were performed according to the manufacturers\u0026rsquo; instructions. Briefly, for MTT assay, at 72h after treatment, the media of the plate was aspirated, after which 50 \u0026micro;L of serum-free media and 50 \u0026micro;L of MTT solution (Sigma-Aldrich; Cat.# M2128) dissolved at 5 mg/mL solution in PBS were added into each well. After incubation at 37\u0026deg;C for 3 hours, 150 \u0026micro;L of MTT solvent (4 mM HCl, 0.1% NP40 in isopropanol) was added into each well. The plate was wrapped in foil and put on an orbital shaker for 15 minutes at room temperature. The plate was read by a BioTek microplate reader at OD\u0026thinsp;=\u0026thinsp;610 nm. Docetaxel IC50 in MTT assay was determined by methods described in the Statistical analysis section below. WST1, CyQuant assay, and Crystal Violet assay were performed as previously described [\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e, \u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e, \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eThe data were assessed using Student's t-test. GraphPad Prism software was used to calculate the statistical significance. The threshold of statistical significance was set at *P\u0026thinsp;\u0026lt;\u0026thinsp;0.05, **P\u0026thinsp;\u0026lt;\u0026thinsp;0.01, ***P\u0026thinsp;\u0026lt;\u0026thinsp;0.001, and ****P\u0026thinsp;\u0026lt;\u0026thinsp;0.0001.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eDetermination of IC50\u003c/h2\u003e \u003cp\u003eThe measured IC50 was calculated using the least square fit of four parameter-sigmoidal curves executed using GraphPad Prism (version 8.4.3) software (GraphPad Software, San Diego, CA, USA, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.graphpad.com/scientific-software/prism/\u003c/span\u003e\u003cspan address=\"https://www.graphpad.com/scientific-software/prism/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e).\u003c/span\u003e Comparison between the measured and predicted values was performed using the correlation coefficient R\u003csup\u003e2\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eDoubling time\u003c/h2\u003e \u003cp\u003eMCF7 cells, PC3 cells, and HELA cells were seeded in triplicate in 96-well plates at 1.00E\u0026thinsp;+\u0026thinsp;4 cells/well. Cell confluency was determined by the SnapCyte\u0026trade; Confluency model at 24h, 48h, 72h and 96h after seeding. Two images were taken per well. Experiments were repeated three times independently and values are expressed in mean. Doubling time (Td) of each cell line in each of the three independent experiments was calculated using the equation below where: Nt is the number of cells at time t; N0 is the number of cells initially at time 0; t is time (days); gr is the growth rate. Number of Cells at Time t (Nt)\u0026thinsp;=\u0026thinsp;N0 * e^(gr * t). Growth Rate (gr) gr\u0026thinsp;=\u0026thinsp;ln(Nt / N0) / t. Doubling Time (Td)\u0026thinsp;=\u0026thinsp;ln(2) / gr.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003cp\u003eN. AN., M.D., and N. F. are co-founders of SnapCyte Solutions Inc. S. A. and JM. S. are employed by SnapCyte Solutions Inc., and C. D., D. G., and JM. P. hold shares in SnapCyte Solutions Inc. The authors declare that these affiliations do not influence the objectivity or integrity of the research presented in this manuscript. All other authors declare no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eConceptualization: N.AN., N.F., C.D., M.D., and D.G.; Methodology: C.P.W., N.AN., N.F., D.G., S.A., N.K.; Investigation: C.P.W., N.AN., S.A., N.F., D.G., N.K., JM.S., and L.R.; Writing \u0026ndash; Original Draft: S.A., N.AN., C.P.W., JM.L., and N.F.; Writing \u0026ndash; Review \u0026amp; Editing: All authors; Funding Acquisition: M.D., N.F., N.AN.; Resources: M.D., C.D., N.F., N.AN., D.G.; Supervision: N.F., D.G., N.AN., M.D.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eWe would like to thank Irina Nelepcu for her technical support, and Dr. Anne Chauchereau for providing IGR-CaP1 cells. This work is supported by the Canadian Institutes of Health Research (CIHR-519572) and MITACS (IT37367).\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eAll data related to the present study are included in the manuscript and supplementary materials. The annotated datasets are protected by University of British Columbia intellectual property and considered trade secrets. Data inquiries can be directed to Dr. Nader Al Nakouzi.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSegeritz, C. P. \u0026amp; Vallier, L. Chap. 9 \u003cem\u003e- Cell Culture: Growing Cells as Model Systems In Vitro\u003c/em\u003e, in Basic Science Methods for Clinical Researchers, M. Jalali, F.Y.L. Saldanha, and M. Jalali, Editors. Academic: Boston. 151\u0026ndash;172. (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGeraghty, R. J. et al. Guidelines for the use of cell lines in biomedical research. \u003cem\u003eBr. J. Cancer\u003c/em\u003e. \u003cb\u003e111\u003c/b\u003e (6), 1021\u0026ndash;1046 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStoddart, M. J. Cell Viability Assays: Introduction, in Mammalian Cell Viability: Methods and Protocols, M.J. Stoddart, Editor. Humana: Totowa, NJ. 1\u0026ndash;6. (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFreshney, R. I. \u003cem\u003eBasic Principles of Cell Culture\u003c/em\u003e, in \u003cem\u003eCulture of Cells for Tissue Engineering\u003c/em\u003e. pp. 1\u0026ndash;22. (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJustus, C. R. et al. \u003cem\u003eIn vitro cell migration and invasion assays\u003c/em\u003e. \u003cem\u003eJ. Vis. Exp.\u003c/em\u003e, \u003cb\u003e2014\u003c/b\u003e(88 ).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRodriguez, E. N. et al. \u003cem\u003eEffect of Seed Cell Density on Specific Growth Rate Using CHO Cells as Model\u003c/em\u003e, in \u003cem\u003eAnimal Cell Technology: From Target to Market: Proceedings of the 17th ESACT Meeting Tyl\u0026ouml;sand, Sweden, June 10\u0026ndash;14\u003c/em\u003e, E. Lindner-Olsson, N. Chatzissavidou, and E. L\u0026uuml;llau, Editors. 2001, Springer Netherlands: Dordrecht. pp. 434\u0026ndash;437. (2001).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGreg\u0026oacute;rio, A. C. et al. Inoculated Cell Density as a Determinant Factor of the Growth Dynamics and Metastatic Efficiency of a Breast Cancer Murine Model. \u003cem\u003ePLoS One\u003c/em\u003e. \u003cb\u003e11\u003c/b\u003e (11), e0165817 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOzturk, S. S. \u0026amp; Palsson, B. O. Effect of initial cell density on hybridoma growth, metabolism, and monoclonal antibody production. \u003cem\u003eJ. Biotechnol.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e (3\u0026ndash;4), 259\u0026ndash;278 (1990).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCocchiararo, I. et al. Back to basics: Optimization of DNA and RNA transfer in muscle cells using recent transfection reagents. \u003cem\u003eExp. Cell. Res.\u003c/em\u003e \u003cb\u003e421\u003c/b\u003e (2), 113392 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMorten, B. C., Scott, R. J. \u0026amp; Avery-Kiejda, K. A. \u003cem\u003eComparison of Three Different Methods for Determining Cell Proliferation in Breast Cancer Cell Lines\u003c/em\u003e. \u003cem\u003eJ. Vis. Exp.\u003c/em\u003e, 2016(115).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiannakeas, V. et al. Analysis of Platelet Count and New Cancer Diagnosis Over a 10-Year Period. \u003cem\u003eJAMA Netw. Open.\u003c/em\u003e \u003cb\u003e5\u003c/b\u003e (1), e2141633 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStella, S. et al. \u003cem\u003eImpact of Different Cell Counting Methods in Molecular Monitoring of Chronic Myeloid Leukemia Patients\u003c/em\u003e. \u003cem\u003eDiagnostics (Basel)\u003c/em\u003e, \u003cb\u003e12\u003c/b\u003e(5). (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMosmann, T. Rapid colorimetric assay for cellular growth and survival: application to proliferation and cytotoxicity assays. \u003cem\u003eJ. Immunol. Methods\u003c/em\u003e. \u003cb\u003e65\u003c/b\u003e (1\u0026ndash;2), 55\u0026ndash;63 (1983).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVembadi, A., Menachery, A. \u0026amp; Qasaimeh, M. A. Cell Cytometry: Review and Perspective on Biotechnological Advances. \u003cem\u003eFront. Bioeng. Biotechnol.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 147 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchroeder, A. B. et al. The ImageJ ecosystem: Open-source software for image visualization, processing, and analysis. \u003cem\u003eProtein Sci.\u003c/em\u003e \u003cb\u003e30\u003c/b\u003e (1), 234\u0026ndash;249 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLanigan, T. M. et al. Real time visualization of cancer cell death, survival and proliferation using fluorochrome-transfected cells in an IncuCyte. \u003cem\u003eJ. Biol. Methods\u003c/em\u003e. \u003cb\u003e7\u003c/b\u003e (2), e133 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavis, J. D. The evolution of the progressive-era hemocytometer. \u003cem\u003eCaduceus\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e (3), 164\u0026ndash;183 (1995).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKuijpers, L. et al. Automated cell counting for Trypan blue-stained cell cultures using machine learning. \u003cem\u003ePLoS One\u003c/em\u003e. \u003cb\u003e18\u003c/b\u003e (11), e0291625 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBiggs, R. \u0026amp; Macmillan, R. L. The Error of the Red Cell Count. \u003cem\u003eJ. Clin. Pathol.\u003c/em\u003e \u003cb\u003e1\u003c/b\u003e (5), 288\u0026ndash;291 (1948).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eManzini, P. et al. Validation of an automated cell counting method for cGMP manufacturing of human induced pluripotent stem cells. \u003cem\u003eBiotechnol. Rep. (Amst)\u003c/em\u003e. \u003cb\u003e33\u003c/b\u003e, e00708 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSarkar, S. et al. Evaluating the quality of a cell counting measurement process via a dilution series experimental design. \u003cem\u003eCytotherapy\u003c/em\u003e \u003cb\u003e19\u003c/b\u003e (12), 1509\u0026ndash;1521 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, M. et al. Improvement of cell counting method for Neubauer counting chamber. \u003cem\u003eJ. Clin. Lab. Anal.\u003c/em\u003e \u003cb\u003e34\u003c/b\u003e (1), e23024 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeoktistova, M., Geserick, P. \u0026amp; Leverkus, M. \u003cem\u003eCrystal Violet Assay for Determining Viability of Cultured Cells\u003c/em\u003e. \u003cem\u003eCold Spring Harb Protoc.\u003c/em\u003e, \u003cb\u003e2016\u003c/b\u003e(4): p. (2016). pdb.prot087379.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerridge, M. V., Herst, P. M. \u0026amp; Tan, A. S. Tetrazolium dyes as tools in cell biology: new insights into their cellular reduction. \u003cem\u003eBiotechnol. Annu. Rev.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 127\u0026ndash;152 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJones, L. J. et al. Sensitive determination of cell number using the CyQUANT cell proliferation assay. \u003cem\u003eJ. Immunol. Methods\u003c/em\u003e. \u003cb\u003e254\u003c/b\u003e (1\u0026ndash;2), 85\u0026ndash;98 (2001).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoleimani, S., Mirzaei, M. \u0026amp; Toncu, D. C. A new method of SC image processing for confluence estimation. \u003cem\u003eMicron\u003c/em\u003e \u003cb\u003e101\u003c/b\u003e, 206\u0026ndash;212 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSunil, B. et al. Opening and closing in morphological image processing. \u003cem\u003eWorld J. Adv. Res. Reviews\u003c/em\u003e. \u003cb\u003e14\u003c/b\u003e (3), 687\u0026ndash;695 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, Z., Zhan, R. \u0026amp; Hu, Y. Automated Confluence Measurement Method for Mesenchymal Stem Cell from Brightfield Microscopic Images. \u003cem\u003eMicrosc. Microanal.\u003c/em\u003e \u003cb\u003e27\u003c/b\u003e (5), 1093\u0026ndash;1101 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAyanzadeh, A. et al. \u003cem\u003eDeep Learning based Segmentation Pipeline for Label-Free Phase-Contrast Microscopy Images\u003c/em\u003e. in \u003cem\u003e2020 28th Signal Processing and Communications Applications Conference (SIU)\u003c/em\u003e. (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang, N. \u0026amp; Yu, F. A Cell Counting Framework Based on Random Forest and Density Map. \u003cem\u003eAppl. Sci.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/app10238346\u003c/span\u003e\u003cspan address=\"10.3390/app10238346\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmidt, U. et al. Cell Detection with Star-Convex Polygons. in Medical Image Computing and Computer Assisted Intervention \u0026ndash; MICCAI 2018. Cham: Springer International Publishing. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStringer, C. et al. Cellpose: a generalist algorithm for cellular segmentation. \u003cem\u003eNat. Methods\u003c/em\u003e. \u003cb\u003e18\u003c/b\u003e (1), 100\u0026ndash;106 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang, N. \u0026amp; Yu, F. A Cell Counting Framework Based on Random Forest and Density Map. \u003cem\u003eAppl. Sci.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (23), 8346 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRonneberger, O., Fischer, P. \u0026amp; Brox, T. \u003cem\u003eU-Net: Convolutional Networks for Biomedical Image Segmentation. in Medical Image Computing and Computer-Assisted Intervention \u0026ndash; MICCAI 2015\u003c/em\u003e (Springer International Publishing, 2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, Z. B., Wang, E. \u0026amp; Zhu, Y. Image segmentation evaluation: a survey of methods. \u003cem\u003eArtif. Intell. Rev.\u003c/em\u003e \u003cb\u003e53\u003c/b\u003e (8), 5637\u0026ndash;5674 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, B. et al. Investigator accuracy: A center-weighted metric for evaluating the location accuracy of image segments in land cover classification. \u003cem\u003eInt. J. Appl. Earth Obs. Geoinf.\u003c/em\u003e \u003cb\u003e122\u003c/b\u003e, 103402 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeliveau, A. et al. Tumorigenicity of the mixed spindle-epithelioid SP6.5 and epithelioid TP17 uveal melanoma cell lines is differentially related to alpha5beta1 integrin expression. \u003cem\u003eInvest. Ophthalmol. Vis. Sci.\u003c/em\u003e \u003cb\u003e42\u003c/b\u003e (12), 3058\u0026ndash;3065 (2001).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKu, J. L. et al. Establishment and characterization of seven human breast cancer cell lines including two triple-negative cell lines. \u003cem\u003eInt. J. Oncol.\u003c/em\u003e \u003cb\u003e43\u003c/b\u003e (6), 2073\u0026ndash;2081 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGagne, J. P. et al. Comparative proteome analysis of human epithelial ovarian cancer. \u003cem\u003eProteome Sci.\u003c/em\u003e \u003cb\u003e5\u003c/b\u003e, 16 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMiura, K. et al. Establishment and characterization of new cell lines of anaplastic pancreatic cancer, which is a rare malignancy: OCUP-A1 and OCUP-A2. \u003cem\u003eBMC Cancer\u003c/em\u003e. \u003cb\u003e16\u003c/b\u003e, 268 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, J. et al. The development of radioresistant oral squamous carcinoma cell lines and identification of radiotherapy-related biomarkers. \u003cem\u003eClin. Transl Oncol.\u003c/em\u003e \u003cb\u003e25\u003c/b\u003e (10), 3006\u0026ndash;3020 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eParveez Ahamed, A. A. et al. In vitro antibacterial activity of MGDG-palmitoyl from Oscillatoria acuminata NTAPC05 against extended-spectrum beta-lactamase producers. \u003cem\u003eJ. Antibiot. (Tokyo)\u003c/em\u003e. \u003cb\u003e70\u003c/b\u003e (6), 754\u0026ndash;762 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePalumbo, P. et al. NOS2 expression in glioma cell lines and glioma primary cell cultures: correlation with neurosphere generation and SOX-2 expression. \u003cem\u003eOncotarget\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e (15), 25582\u0026ndash;25598 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Vries, G. H. \u0026amp; Boullerne, A. I. Glial cell lines: an overview. \u003cem\u003eNeurochem Res.\u003c/em\u003e \u003cb\u003e35\u003c/b\u003e (12), 1978\u0026ndash;2000 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlemi, M. et al. Anti-inflammatory effect of seeds and callus of Nigella sativa L. extracts on mix glial cells with regard to their thymoquinone content. \u003cem\u003eAAPS PharmSciTech\u003c/em\u003e. \u003cb\u003e14\u003c/b\u003e (1), 160\u0026ndash;167 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHolliday, D. L. \u0026amp; Speirs, V. Choosing the right cell line for breast cancer research. \u003cem\u003eBreast Cancer Res.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e (4), 215 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, H. Y. et al. Transforming growth factor-beta1-induced epithelial-mesenchymal transition in human esophageal squamous cell carcinoma via the PTEN/PI3K signaling pathway. \u003cem\u003eOncol. Rep.\u003c/em\u003e \u003cb\u003e32\u003c/b\u003e (5), 2134\u0026ndash;2142 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKiss, A. et al. MTT Test and Time-lapse Microscopy to Evaluate the Antitumor Potential of Nucleoside Analogues. \u003cem\u003eAnticancer Res.\u003c/em\u003e \u003cb\u003e41\u003c/b\u003e (1), 137\u0026ndash;149 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNakamura, R. et al. Establishment and characterization of a human lymphatic endothelial cell line. \u003cem\u003eBiol. Pharm. Bull.\u003c/em\u003e \u003cb\u003e37\u003c/b\u003e (4), 683\u0026ndash;687 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThonhoff, J. R. et al. Compatibility of human fetal neural stem cells with hydrogel biomaterials in vitro. \u003cem\u003eBrain Res.\u003c/em\u003e \u003cb\u003e1187\u003c/b\u003e, 42\u0026ndash;51 (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChatree, K. et al. \u003cem\u003eDistinctions in bone matrix nanostructure, composition, and formation between osteoblast-like cells, MG-63, and human mesenchymal stem cells, UE7T-13.\u003c/em\u003e Heliyon, 9(5): p. e15556. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaberlein, K. et al. \u003cem\u003eGenetic Characterization of Rat Hepatic Stellate Cell Line PAV-1\u003c/em\u003e. \u003cem\u003eCells\u003c/em\u003e, \u003cb\u003e12\u003c/b\u003e(12). (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShepherd, T. G. et al. Corrigendum: Primary culture of ovarian surface epithelial cells and ascites-derived ovarian cancer cells from patients. \u003cem\u003eNat. Protoc.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (9), 1457 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLobsenz, A. \u0026amp; Seidler, P. \u003cem\u003eThe Confluence Analysis Program: A User-Friendly Tool for Automated Cell Confluence Measurement and Visualization\u003c/em\u003e, in \u003cem\u003ePreprints\u003c/em\u003e. Preprints. (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKang, C. et al. Variability Matters: Evaluating Inter-Rater Variability in Histopathology for Robust Cell Detection. in Computer Vision \u0026ndash; ECCV 2022 Workshops. Cham: Springer Nature Switzerland. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStringer, C. \u0026amp; Pachitariu, M. \u003cem\u003eCellpose3: one-click image restoration for improved cellular segmentation.\u003c/em\u003e bioRxiv, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCutler, K. J. et al. Omnipose: a high-precision morphology-independent solution for bacterial cell segmentation. \u003cem\u003eNat. Methods\u003c/em\u003e. \u003cb\u003e19\u003c/b\u003e (11), 1438\u0026ndash;1448 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim, S. I. et al. Application of a non-hazardous vital dye for cell counting with automated cell counters. \u003cem\u003eAnal. Biochem.\u003c/em\u003e \u003cb\u003e492\u003c/b\u003e, 8\u0026ndash;12 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeo, D. et al. \u003cem\u003eA Field-Portable Cell Analyzer without a Microscope and Reagents\u003c/em\u003e. \u003cem\u003eSens. (Basel)\u003c/em\u003e, \u003cb\u003e18\u003c/b\u003e(1). (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Noorden, R. More than 10,000 research papers were retracted in 2023 - a new record. \u003cem\u003eNature\u003c/em\u003e \u003cb\u003e624\u003c/b\u003e (7992), 479\u0026ndash;481 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChiu, C. H. et al. Systematic Quantification of Cell Confluence in Human Normal Oral Fibroblasts. \u003cem\u003eAppl. Sci.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/app10249146\u003c/span\u003e\u003cspan address=\"10.3390/app10249146\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Valen, D. A. et al. Deep Learning Automates the Quantitative Analysis of Individual Cells in Live-Cell Imaging Experiments. \u003cem\u003ePLoS Comput. Biol.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e (11), e1005177 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcQuin, C. et al. CellProfiler 3.0: Next-generation image processing for biology. \u003cem\u003ePLoS Biol.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e (7), e2005970 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, L. C. et al. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. \u003cem\u003eIEEE Trans. Pattern Anal. Mach. Intell.\u003c/em\u003e \u003cb\u003e40\u003c/b\u003e (4), 834\u0026ndash;848 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSkeltved, N. et al. Bispecific T cell-engager targeting oncofetal chondroitin sulfate induces complete tumor regression and protective immune memory in mice. \u003cem\u003eJ. Exp. Clin. Cancer Res.\u003c/em\u003e \u003cb\u003e42\u003c/b\u003e (1), 106 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChauchereau, A. et al. Stemness markers characterize IGR-CaP1, a new cell line derived from primary epithelial prostate cancer. \u003cem\u003eExp. Cell. Res.\u003c/em\u003e \u003cb\u003e317\u003c/b\u003e (3), 262\u0026ndash;275 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAl Nakouzi, N. et al. Cabazitaxel Remains Active in Patients Progressing After Docetaxel Followed by Novel Androgen Receptor Pathway Targeted Therapies. \u003cem\u003eEur. Urol.\u003c/em\u003e \u003cb\u003e68\u003c/b\u003e (2), 228\u0026ndash;235 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeacock, J. W. et al. SEMA3C drives cancer growth by transactivating multiple receptor tyrosine kinases via Plexin B1. \u003cem\u003eEMBO Mol. Med.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (2), 219\u0026ndash;238 (2018).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6280571/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6280571/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePrecise assessment of cell growth, count, and viability is crucial in biological and medical research. Traditional cell analytics involve manual processes, such as cell counting or reagent-based approaches that are user-dependent and prone to bias. Semi-automated systems for counting cells, tracking cell growth, and determining viability, have been introduced over the past decades. However, these methods are often time-consuming, require labeling steps, and involve costly instrumentation and consumables. Changes in cell growth and/or viability create biological patterns that can be interpreted by artificial intelligence (AI). Here, we report the development and validation of SnapCyte\u0026trade;, an AI application that performs accurate, unbiased, label- and reagent-free cell analyses from basic cell culture images. Using cell lines with diverse morphologies in various culture conditions, we generated a comprehensive and fully annotated image database that was used for AI education. Convolutional neural networks were employed for cell localization and iterative training loops until a stable performance of \u0026gt;\u0026thinsp;95% accuracy was obtained for all readouts. The fully trained AI demonstrated high Precision and Recall and performed with greater accuracy and less variation as compared to standard methods. As the SnapCyte\u0026trade; analyses are performed on cell images only, data acquisition is non-invasive to the experimental setup, enabling real-time use of cells in downstream assays. In summary, SnapCyte\u0026trade; is a fast and accurate cell analytics platform, resistant to user variations and independent of reagents or specific equipment, with improved performance over current cell analytics methodologies.\u003c/p\u003e","manuscriptTitle":"Enhanced Precision in Cell Culture Analytics: Leveraging Artificial Intelligence for Unbiased and Non-Destructive Assessment of Cell Growth and Viability.","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-10 08:07:27","doi":"10.21203/rs.3.rs-6280571/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ffd252d0-7457-496c-938c-a441aba378b4","owner":[],"postedDate":"April 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":46847215,"name":"Biological sciences/Biotechnology"},{"id":46847216,"name":"Biological sciences/Cell biology"}],"tags":[],"updatedAt":"2025-09-19T10:16:25+00:00","versionOfRecord":[],"versionCreatedAt":"2025-04-10 08:07:27","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6280571","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6280571","identity":"rs-6280571","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}