A Quantitative Approach to Optimise Microcarrier Surface Area for Smarter Adherent Cell Culture

A Quantitative Approach to Optimise Microcarrier Surface Area for Smarter Adherent Cell Culture | 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 Short Report A Quantitative Approach to Optimise Microcarrier Surface Area for Smarter Adherent Cell Culture Rayan Metha, Ashley Chen, Jane Toting, Sarina Ghobadi, Payar Radfar, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9391084/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract Optimising adherent cell culture requires precise microcarrier quantification to transition from traditional mass-based (g/L) reporting to more rational surface-area-dependent process control. Current monitoring workflows are heavily hindered by human error, subjectivity, and the time-intensive nature of manual counting. The objective of this research is to provide a highly efficient, accurate and accessible technological solution that streamlines particle counting and analysis, therefore significantly contributing to the optimisation of biomanufacturing workflows. We present an adaptive, automated computer vision framework utilising the Gradient Hough Transform (GHT) integrated with a Bayesian Optimisation (BO) loop for rapid, supervised parameter tuning. For aggregated microcarrier images, the system achieves a median Absolute Percentage Error (APE) of 1.2% for counts, 1.7% for average diameter estimation, and 7.8% for population variation estimation. In these conditions, the system outperforms both manual expert benchmarks and semi-automated ImageJ plugins across every metric. While performance declines on noisy cellular images, the model's 24.7% count error still significantly outperforms the 47.2% error produced by ImageJ while remaining significantly faster than manual methods. Overall, the framework provides a rapid and reproducible alternative for microcarrier enumeration and characterisation, reducing user time from several minutes to seconds per image. While identifying "critical boundaries" in noisy cellular images, this work establishes a robust, user-friendly baseline for automated microcarrier characterisation in regenerative medicine and vaccine production. Microcarrier Bioprocessing Adherent cell culture Vaccine production Figures Figure 1 Figure 2 Figure 3 Introduction Accurate microcarrier quantification is essential for optimising large-scale cell culture systems used in regenerative medicine, vaccine production, and cultivated-meat manufacturing. In these bioprocesses, microcarriers provide the physical surface on which adherent cells attach and proliferate [ 1 , 2 ]. Despite this surface-dependent growth mechanism, microcarrier dosing is often reported and controlled by mass concentration in grams per litre. This practice relies on the assumption that weight directly correlates with available surface area [ 3 ] [ 4 ], which is a simplification that fails when microcarriers differ in size, density, or porosity. Consequently, bioprocesses optimised by weight rather than surface area can suffer from poor reproducibility and inefficient media utilisation. Transitioning to the accurate determination of particle number and size distribution is critical because knowing the total exposed surface area enables rational optimisation of seeding density and nutrient delivery [ 1 ]. However, existing methods for counting microcarriers remain limited and represent a primary bottleneck in biomanufacturing. Manual counting under a microscope is labour-intensive, error-prone, and subjective, while commercial image-analysis systems are costly and often require dedicated hardware or complex software environments, such as high-content imaging platforms or custom pipelines in ImageJ/Fiji or CellProfiler. Furthermore, subtle cognitive biases often lead users to preferentially count particles that appear more distinct or uniform in size, which introduces a significant and systematic human error into the results [ 5 ] [ 6 ]. Computer vision has emerged as a powerful paradigm to overcome these inaccuracies by objectively extracting quantitative information based on visual data [ 7 ]. Specific stand-alone computer vision applications for particle detection and distribution analysis may require advanced programming and tend to use dedicated hardware and proprietary software [ 1 ] [ 5 ]. Furthermore, while semi-automated particle counting software such as the ParticleAnalyzer (ImageJ API) plugin does reduce labour, the process remains time-consuming compared to fully integrated pipelines, as images often require manual adjustment of parameters such as thresholding, diameter ranges, cell circularity and contrast [ 6 ]. The process is faster than manual counting but can still be time-consuming overall. Time is saved on per-image counting, but substantial effort is often required for preprocessing, quality control and per-dataset parameter selection, steps that may offset raw time saving on smaller datasets. Moreover, despite parameter tuning, a degree of imprecision persists due to segmentation variability and image noise. The limitation of these approaches underscores a pressing need for more reliable, automated and user-friendly solutions that can maintain pace with the increasing demands of modern bioprocessing. Recently, deep learning methods such as Convolutional Neural Networks (CNNs) have been applied to scientific particle detection and classification. These models, equipped with often hundreds of thousands of parameters, are trained from scratch using a vast generalised image dataset and then fine-tuned for specific applications, allowing high degrees of accuracy in object detection and image parsing [ 8 ]. While deep learning approaches can be more robust to variations in particle morphology, occlusion, and contrast [ 9 ], they typically require large, annotated datasets [ 10 ] and substantial computational resources for training and deployment. Consequently, they are most suitable when sufficient ground-truth data can be assembled or when high robustness across diverse imaging conditions is required. This frequently involves the labelling of hundreds to thousands of images [ 8 ], often providing a significant bottleneck for specialised applications where such data is scarce. Traditional computer vision methods for object detection rely on explicitly defined mathematical models in combination with heuristic methods. Among these, Hough Transform stands out as a foundational method for detecting geometric shapes like lines, circles and ellipses within images, particularly due to its merits in noise immunity and expandability [ 11 ]. Specifically, the Hough Circle Transform is widely used for identifying circular patterns by transforming points into a parameter space where circles are represented as intersections or accumulations of votes [ 12 ]. These traditional image processing approaches provide interpretable and efficient solutions that perform well when particle morphology and illumination are relatively consistent. Their rule-based nature allows fine control and transparency in the detection process, making them suitable for domains where physical interpretability and low data demands are priorities. In this work, we present a robust automated system based on morphological segmentation that bridges the gap between traditional computer vision and data-driven adaptation. Built around the Gradient Hough Transform (GHT) [ 11 ] [ 12 ], this pipeline leverages a supervised learning mechanism implemented with Bayesian Optimisation (BO) [ 13 ]. This framework provides a unique training capability that allows users to automatically tune detection parameters using only a few labelled images. Results demonstrate that this system achieves a median Absolute Percentage Error (APE) of 1.5% or less on microcarrier images while performing analysis over 300 times faster than manual methods. This technology delivers a significant impact by supporting the scalable monitoring and optimisation of biomanufacturing through rapid and reproducible microcarrier characterisation. Materials and Methods 1. Model Pipeline The methodology for analysing particles in an image is subject to a set of parameters that are optimised via BO for each image class to ensure a robust solution. The core of the model is built around the GHT. The Hough Transform operates through a two-step process that begins with edge detection via the Canny algorithm, followed by circle detection using the Hough Gradient Method. During the initial stage, the system identifies candidate edge pixels by applying Gaussian blurs and Sobel filters to determine gradient orientations, which are then refined through non-maximum suppression and hysteresis thresholding. These refined edge points then participate in a "voting" system where they cast votes for potential circle centres along lines perpendicular to their gradient directions. Finally, for candidate centres that exceed a predefined vote threshold, the algorithm calculates the corresponding radius to define the specific parameters of the detected circular particles. The automated particle counting system is based on a morphological segmentation pipeline built around the GHT [ 12 ] implemented via the OpenCV library [ 14 ]. As shown in Fig. 1 , images are first converted to grayscale and blurred using predefined settings. The images are then processed using GHT, followed by intensity- and diameter-based thresholding using the same pipeline parameters. To overcome the rigidity of traditional computer vision models, the GHT is embedded within a supervised BO framework using Gaussian Processes [ 13 ]. This framework provides a unique training capability that allows the system to learn optimal pipeline parameters, including blur levels, intensity thresholds, and accumulator sensitivities, from a small set of user-provided labelled images. For a given image class, the BO mechanism explores the parameter space over 200 evaluations, including40 random initialisations, to identify a stable configuration that generalises to unseen data. The utility function for this optimisation is a custom efficiency metric (E), which combines the F1-score for detection fidelity with the Normalised Average Diameter Error (NAE) for geometric precision \(\:E={F}_{1}\cdot\:(1-NAE)\) . An E value of 1 represents a perfect annotator. All experiments were conducted in Python version 3.13 using HoughCircles from OpenCV version 4.11.0 [ 14 ] and gp_minimize (BO) from scikit-optimize version 0.10.2 [ 15 ]. 2. Data Acquisition and Image Classes A total of six image classes (A–F) (representative images for each class are shown in supplementary table 1 ), each containing five images, were used to benchmark the model across a broad range of microscopy conditions. Classes A–C consist of microcarriers characterised by high contrast and spherical morphology, while classes D–F consist of cellular images with higher background noise and irregular shapes. Ground truth annotations were generated using the Computer Vision Annotation Tool (CVAT), where an expert manually marked the precise centre coordinates and diameters for every visible particle to establish the true population characteristics. 3. Experimental Benchmarking The performance of the trained automated system was rigorously evaluated using Leave-One-Out Cross-Validation (LOOCV) within each image class to ensure that the stable parameter configurations identified by the Bayesian framework could effectively generalise to unseen data [ 13 ]. To provide a comprehensive assessment, the model was benchmarked against two standard practices that currently define the industry baseline. The first of these was a semi-automated analysis where an expert spent 30 minutes per image to manually determine the optimal binarisation and thresholding configurations using the ImageJ “Analyze Particles” plugin, version 1.54 [ 16 ]. This process highlights the significant time investment and iterative nature of traditional tools which this study seeks to streamline. The second benchmark replicated common industry quality control practices through a manual expert assessment; in this case, an experienced lab technician produced a total manual count for each image and measured the diameters of 20 randomly selected particles to derive population-level estimates for the mean diameter and coefficient of variation. By comparing the automated pipeline against these manual and semi-automated workflows, we aimed to quantify the reduction in human error and the potential impact of the technology on bioprocess monitoring. All three methods were compared using a full ground-truth dataset generated via the CVAT, where an expert manually marked the precise centre coordinates and diameters of every visible particle across all six image classes. The primary metrics for comparison were the median APE in particle counts, the average diameter, and the Coefficient of Variation (CV) of the diameters. These specific metrics were selected to evaluate the system’s capacity to provide the rapid, high-throughput data required for a transition from mass-based dosing to rational surface-area-dependent process control. Detailed mathematical definitions for the intersection-over-union (IOU) thresholds used to classify detections, alongside specific error calculations, are provided in the Supplementary Information. Results and Discussion The automated system’s performance was evaluated by aggregating results into two meta-groups: Microcarriers (Classes A–C) and Cellular (Classes D–F) (Supplementary table 1 ). The primary objective was to demonstrate a high-throughput alternative that eliminates human error and provides the geometric precision required for surface-area-based process control. 1. Performance on Microcarrier Populations For microcarrier images, the model achieved exceptional accuracy, outperforming both manual benchmarks and semi-automated ImageJ plugins across all primary metrics. As shown in Fig. 2 a, the model recorded a median count APE of 1.2%, compared to 1.6% for manual counts and 2.2% for ImageJ. Geometric precision was similarly high, with a median diameter APE of 1.7% (Fig. 2 b) and a CV error of 7.8% (Fig. 2 c). The success of the system in these conditions validates the strength of the BO framework. By exploring the parameter space over 200 evaluations, the BO mechanism identified stable configurations that generalised effectively to unseen high-contrast images, positioning the system as an immediate practical replacement for an accurate and automated microcarrier particle counting. Crucially, as visualised in Fig. 2 d, the system completed these analyses over 300x faster than manual methods, reducing user time from approximately ten minutes to just seconds per image, delivering a substantial high-throughput impact. 2. Critical Boundaries and Failure Modes While the system provides a robust solution for microcarriers, performance declined in noisy cellular environments, with a median count APE of 24.7% (Fig. 2 a). This decline defines the "critical boundaries" of the GHT approach. To guide future development, we identified four distinct failure modes (illustrated in Fig. 3 ): Low Particle-Background Contrast: As shown in Fig. 3 A, low-contrast images (e.g., Class D) produce few distinct boundary gradients. Since GHT detectors rely on consistent gradient vectors to cast "votes" for circle centres, low contrast results in insufficient voter accumulation and significant undercounting [ 14 ]. Non-Particle Artefacts: Cellular images often contain smudges or debris (Fig. 3 B). These spurious edges cause false peaks in the accumulator space, leading to false-positive detections, while simultaneously obscuring the edges of real objects [ 17 ]. Heterogeneous Colour and Size: When particles exhibit diverse colours or a wide size range (Fig. 3 C), the BO mechanism may struggle to identify a single parameter interval that captures all instances without overgeneralising [ 13 ]. High Degree of Clustering: Dense regions, such as the outlier image 2 in Class A, proved challenging (Fig. 3 D). When particles overlap, their edge points cast votes that merge into a single broad peak rather than distinct maxima. The "minDist" parameter, designed to prevent duplicate detections, unfortunately suppresses true detections when particles are closer than the specified threshold, leading to systematic undercounting in clustered regions [ 14 ]. 3. Discussion of Impact The industry’s current reliance on mass-based dosing (g/L) is a simplification that fails to account for critical variations in particle size and density, often leading to poor reproducibility and sub-optimal cell attachment. Transitioning to surface-area metrics enables a more rational process design, allowing for the systematic optimisation of seeding densities and nutrient delivery. While manual experts can delineate indistinct boundaries, they remain limited by a ten-minute-per-sample processing time and are prone to systematic cognitive biases that introduce significant human error. This system provides an objective, reproducible baseline with a significant impact, operating 300x faster than manual methods. Although performance on complex cellular morphologies remains a challenge, the model’s 24.7% count error still significantly outperforms the 47.2% error produced by standard ImageJ thresholding. This establishes adaptive computer vision as a superior tool for bioprocess monitoring where speed and objectivity are paramount. Conclusion This study successfully validated an adaptive computer vision framework that provides high precision circular particle quantification while eliminating the subjectivity and human error inherent in labour-intensive manual particle counting. By achieving a median count APE of 1.2% and performing analysis over 300x faster than experts, the system enables a critical transition toward surface-area-based process control. While complex cellular morphologies define the current "critical boundaries" of the GHT, this work delivers an immediate, high-throughput impact for routine microcarrier characterisation and establishes a reproducible baseline for future deep-learning-based bioprocess monitoring. Declarations The authors have no conflicts of interest to declare that are relevant to the content of this article. Acknowledgement We thank Professor Mohsen Asadnia for his support and valuable discussions. The authors acknowledge funding support from the NSW Government TechVoucher Program (Project ID: TV2500010). This project was delivered in collaboration with Macquarie University and Smarty MCs Pty Ltd. References Cox TR et al (2025) Maximising adherent cell production via customisable and dissolvable bio-polymer microcarriers. Biomed Mater, 20(5) Ding L et al (2023) Scaling up stem cell production: harnessing the potential of microfluidic devices. Biotechnol Adv 69:108271 Levine DW, Wang DIC, Thilly WG (2004) Optimization of growth surface parameters in microcarrier cell culture. Biotechnol Bioeng 21(5):821–845 Croughan MS, Hamel JF, Wang DI (1988) Effects of microcarrier concentration in animal cell culture. Biotechnol Bioeng 32(8):975–982 Gray AJ et al (2002) Cell identification and sizing using digital image analysis for estimation of cell biomass in High Rate Algal Ponds. J Appl Phycol 14(3):193–204 Igathinathane C et al (2008) Shape identification and particles size distribution from basic shape parameters using ImageJ. Comput Electron Agric 63(2):168–182 Zhang J et al (2022) A comprehensive review of image analysis methods for microorganism counting: from classical image processing to deep learning approaches. Artif Intell Rev 55(4):2875–2944 Mahony NO et al (2020) Deep Learning vs. Traditional Computer Vision . Vol. 943 Li Z et al (2022) A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects. IEEE Trans Neural Netw Learn Syst 33(12):6999–7019 Yu Z, Bajaj C (2004) Detecting circular and rectangular particles based on geometric feature detection in electron micrographs. J Struct Biol 145(1–2):168–180 Hassanein AS et al (2015) A Survey on Hough Transform, Theory, Techniques and Applications. Yuen HK et al (1990) Comparative study of Hough Transform methods for circle finding. Image Vis Comput 8(1):71–77 Shahriari B et al (2016) Taking the Human Out of the Loop: A Review of Bayesian Optimization. Proceedings of the IEEE, 104(1): pp. 148–175 Bradski G (2025) OpenCV 4.x – Image Processing Module: Feature Detection (imgproc). OpenCV Developers contributors, s.-o., skopt.gp_minimize Developers I ImageJ - Particle Analysis Gyeongyong H, Hun C, Jihong K (2016) Robust Hough Transform with Edge Strength. Int J Eng Res and, V5(11). Additional Declarations No competing interests reported. Supplementary Files ParticleCounterSupplementalInfo.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 26 Apr, 2026 Editor assigned by journal 16 Apr, 2026 Submission checks completed at journal 16 Apr, 2026 First submitted to journal 11 Apr, 2026 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-9391084","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Short Report","associatedPublications":[],"authors":[{"id":629991508,"identity":"74ddb493-2c6b-41ee-95e2-31dd3a42540f","order_by":0,"name":"Rayan Metha","email":"","orcid":"","institution":"Smart MCs PTY LTD","correspondingAuthor":false,"prefix":"","firstName":"Rayan","middleName":"","lastName":"Metha","suffix":""},{"id":629991523,"identity":"8325e54a-8353-42ac-acf1-6df890f8b2ab","order_by":1,"name":"Ashley Chen","email":"","orcid":"","institution":"Smart MCs PTY LTD","correspondingAuthor":false,"prefix":"","firstName":"Ashley","middleName":"","lastName":"Chen","suffix":""},{"id":629991526,"identity":"5c18d1a0-ca8a-4b9b-a8bf-11e995b76ed2","order_by":2,"name":"Jane Toting","email":"","orcid":"","institution":"Smart MCs PTY LTD","correspondingAuthor":false,"prefix":"","firstName":"Jane","middleName":"","lastName":"Toting","suffix":""},{"id":629991538,"identity":"0ce8f443-a15f-411b-be65-9d8cd2298453","order_by":3,"name":"Sarina Ghobadi","email":"","orcid":"","institution":"Smart MCs PTY LTD","correspondingAuthor":false,"prefix":"","firstName":"Sarina","middleName":"","lastName":"Ghobadi","suffix":""},{"id":629991551,"identity":"774278bd-bff1-4a16-b40a-8a320c27a74e","order_by":4,"name":"Payar Radfar","email":"","orcid":"","institution":"Smart MCs PTY LTD","correspondingAuthor":false,"prefix":"","firstName":"Payar","middleName":"","lastName":"Radfar","suffix":""},{"id":629991553,"identity":"b3fb7bef-c4d5-4f6f-b8b2-7239d5f7ce04","order_by":5,"name":"Majid Ebrahimi Warkiani","email":"","orcid":"","institution":"Smart MCs PTY LTD","correspondingAuthor":false,"prefix":"","firstName":"Majid","middleName":"Ebrahimi","lastName":"Warkiani","suffix":""},{"id":629991561,"identity":"4523f812-7993-4e7a-9abd-0605b153c3be","order_by":6,"name":"Lin Ding","email":"","orcid":"","institution":"Smart MCs PTY LTD","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Ding","suffix":""},{"id":629991563,"identity":"cb1a034e-3d11-4fe7-9185-66d1d668ccd1","order_by":7,"name":"Rouzbeh Abbassi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABC0lEQVRIiWNgGAWjYJACZgaGAwwMEjwgtoQcWIiHFC3GJGthSGwgpMW8vcf4c0HFHQb+2b0HP/74Y5G+tj2B8cHbNgZ5gwPYtcicOWNgPOPMMwaJO+eSpXnbJHK3nXnAbDi3jcFwAw4tEhI5Bsm8bYcZGG7kGEgzNgC13EhgA+plYMSn5TDvv8MM8jdyjH/++CORbnYjgf03UIs9Hi2GzbwNhxkMbuSYSfCwSSQAtbAxA7Uk4tTCc6yYecaxwzyGd86lWQP9YrjtzMNmyTnnJJJn4tLC3rz5c0HNYTm5272Hb/74Uydvdjz54Ic3ZTa2fTi0wAByRICjRgK/ejSQQJLqUTAKRsEoGP4AANBPXYQeXurrAAAAAElFTkSuQmCC","orcid":"","institution":"Macquarie University","correspondingAuthor":true,"prefix":"","firstName":"Rouzbeh","middleName":"","lastName":"Abbassi","suffix":""}],"badges":[],"createdAt":"2026-04-12 01:54:02","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9391084/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9391084/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108839514,"identity":"2f6d5295-a9bd-4f37-b8d0-de099d057cce","added_by":"auto","created_at":"2026-05-09 00:47:39","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":119810,"visible":true,"origin":"","legend":"\u003cp\u003eBlock diagram of the automated particle counting pipeline showing the integration of BO for supervised parameter tuning and the resulting data-driven adaptation of detection parameters\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9391084/v1/c737e0cd1202145c001689f1.png"},{"id":108839515,"identity":"673f9ca0-8dc9-4b2e-9f0a-f765cb166506","added_by":"auto","created_at":"2026-05-09 00:47:39","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":385130,"visible":true,"origin":"","legend":"\u003cp\u003eComparative performance of the model, ImageJ, and manual expert methods across microcarrier and cellular image groups showing APE for (a) total particle count, (b) average diameter estimate, and (c) coefficient of variation of diameters estimate. (d) shows the relative speeds of each method.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9391084/v1/90d66eebfd5cd6864d70a632.png"},{"id":108839513,"identity":"2a3d34be-786c-4723-8005-e42be854dea2","added_by":"auto","created_at":"2026-05-09 00:47:39","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":209485,"visible":true,"origin":"","legend":"\u003cp\u003eQualitative assessment of the four primary failure modes for the GHT model including (a) the difference between the high particle-background contrast in microcarrier images (left) to the low contrast in cellular images (right), (b) the presence of non-particle artefacts in cellular images, (c) detection instability with heterogeneous particle colours, and (d) systematic undercounting in high-clustering regions (left) as opposed to low-clustering regions (right)\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9391084/v1/583f19945ffb35d6e13afb64.png"},{"id":108839516,"identity":"22a44e1d-3265-4b80-b5e1-a5674398718c","added_by":"auto","created_at":"2026-05-09 00:47:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":882261,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9391084/v1/d6bf6f74-c739-42ef-9aeb-c254f48daf28.pdf"},{"id":108839512,"identity":"d0ba84f5-6038-422d-bb15-5a24738a006b","added_by":"auto","created_at":"2026-05-09 00:47:39","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":606473,"visible":true,"origin":"","legend":"","description":"","filename":"ParticleCounterSupplementalInfo.docx","url":"https://assets-eu.researchsquare.com/files/rs-9391084/v1/c5c6c458f95f04e02da5c6e7.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Quantitative Approach to Optimise Microcarrier Surface Area for Smarter Adherent Cell Culture","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAccurate microcarrier quantification is essential for optimising large-scale cell culture systems used in regenerative medicine, vaccine production, and cultivated-meat manufacturing. In these bioprocesses, microcarriers provide the physical surface on which adherent cells attach and proliferate [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Despite this surface-dependent growth mechanism, microcarrier dosing is often reported and controlled by mass concentration in grams per litre. This practice relies on the assumption that weight directly correlates with available surface area [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], which is a simplification that fails when microcarriers differ in size, density, or porosity. Consequently, bioprocesses optimised by weight rather than surface area can suffer from poor reproducibility and inefficient media utilisation. Transitioning to the accurate determination of particle number and size distribution is critical because knowing the total exposed surface area enables rational optimisation of seeding density and nutrient delivery [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eHowever, existing methods for counting microcarriers remain limited and represent a primary bottleneck in biomanufacturing. Manual counting under a microscope is labour-intensive, error-prone, and subjective, while commercial image-analysis systems are costly and often require dedicated hardware or complex software environments, such as high-content imaging platforms or custom pipelines in ImageJ/Fiji or CellProfiler. Furthermore, subtle cognitive biases often lead users to preferentially count particles that appear more distinct or uniform in size, which introduces a significant and systematic human error into the results [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Computer vision has emerged as a powerful paradigm to overcome these inaccuracies by objectively extracting quantitative information based on visual data [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSpecific stand-alone computer vision applications for particle detection and distribution analysis may require advanced programming and tend to use dedicated hardware and proprietary software [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Furthermore, while semi-automated particle counting software such as the ParticleAnalyzer (ImageJ API) plugin does reduce labour, the process remains time-consuming compared to fully integrated pipelines, as images often require manual adjustment of parameters such as thresholding, diameter ranges, cell circularity and contrast [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. The process is faster than manual counting but can still be time-consuming overall. Time is saved on per-image counting, but substantial effort is often required for preprocessing, quality control and per-dataset parameter selection, steps that may offset raw time saving on smaller datasets. Moreover, despite parameter tuning, a degree of imprecision persists due to segmentation variability and image noise. The limitation of these approaches underscores a pressing need for more reliable, automated and user-friendly solutions that can maintain pace with the increasing demands of modern bioprocessing.\u003c/p\u003e \u003cp\u003eRecently, deep learning methods such as Convolutional Neural Networks (CNNs) have been applied to scientific particle detection and classification. These models, equipped with often hundreds of thousands of parameters, are trained from scratch using a vast generalised image dataset and then fine-tuned for specific applications, allowing high degrees of accuracy in object detection and image parsing [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. While deep learning approaches can be more robust to variations in particle morphology, occlusion, and contrast [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], they typically require large, annotated datasets [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] and substantial computational resources for training and deployment. Consequently, they are most suitable when sufficient ground-truth data can be assembled or when high robustness across diverse imaging conditions is required. This frequently involves the labelling of hundreds to thousands of images [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], often providing a significant bottleneck for specialised applications where such data is scarce.\u003c/p\u003e \u003cp\u003eTraditional computer vision methods for object detection rely on explicitly defined mathematical models in combination with heuristic methods. Among these, Hough Transform stands out as a foundational method for detecting geometric shapes like lines, circles and ellipses within images, particularly due to its merits in noise immunity and expandability [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Specifically, the Hough Circle Transform is widely used for identifying circular patterns by transforming points into a parameter space where circles are represented as intersections or accumulations of votes [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. These traditional image processing approaches provide interpretable and efficient solutions that perform well when particle morphology and illumination are relatively consistent. Their rule-based nature allows fine control and transparency in the detection process, making them suitable for domains where physical interpretability and low data demands are priorities.\u003c/p\u003e \u003cp\u003eIn this work, we present a robust automated system based on morphological segmentation that bridges the gap between traditional computer vision and data-driven adaptation. Built around the Gradient Hough Transform (GHT) [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], this pipeline leverages a supervised learning mechanism implemented with Bayesian Optimisation (BO) [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. This framework provides a unique training capability that allows users to automatically tune detection parameters using only a few labelled images. Results demonstrate that this system achieves a median Absolute Percentage Error (APE) of 1.5% or less on microcarrier images while performing analysis over 300 times faster than manual methods. This technology delivers a significant impact by supporting the scalable monitoring and optimisation of biomanufacturing through rapid and reproducible microcarrier characterisation.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\n\u003ch3\u003e1. Model Pipeline\u003c/h3\u003e\n\u003cp\u003eThe methodology for analysing particles in an image is subject to a set of parameters that are optimised via BO for each image class to ensure a robust solution. The core of the model is built around the GHT. The Hough Transform operates through a two-step process that begins with edge detection via the Canny algorithm, followed by circle detection using the Hough Gradient Method. During the initial stage, the system identifies candidate edge pixels by applying Gaussian blurs and Sobel filters to determine gradient orientations, which are then refined through non-maximum suppression and hysteresis thresholding. These refined edge points then participate in a \"voting\" system where they cast votes for potential circle centres along lines perpendicular to their gradient directions. Finally, for candidate centres that exceed a predefined vote threshold, the algorithm calculates the corresponding radius to define the specific parameters of the detected circular particles.\u003c/p\u003e \u003cp\u003eThe automated particle counting system is based on a morphological segmentation pipeline built around the GHT [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] implemented via the OpenCV library [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, images are first converted to grayscale and blurred using predefined settings. The images are then processed using GHT, followed by intensity- and diameter-based thresholding using the same pipeline parameters.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo overcome the rigidity of traditional computer vision models, the GHT is embedded within a supervised BO framework using Gaussian Processes [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. This framework provides a unique training capability that allows the system to learn optimal pipeline parameters, including blur levels, intensity thresholds, and accumulator sensitivities, from a small set of user-provided labelled images.\u003c/p\u003e \u003cp\u003eFor a given image class, the BO mechanism explores the parameter space over 200 evaluations, including40 random initialisations, to identify a stable configuration that generalises to unseen data. The utility function for this optimisation is a custom efficiency metric (E), which combines the F1-score for detection fidelity with the Normalised Average Diameter Error (NAE) for geometric precision\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:E={F}_{1}\\cdot\\:(1-NAE)\\)\u003c/span\u003e \u003c/span\u003e. An E value of 1 represents a perfect annotator.\u003c/p\u003e \u003cp\u003eAll experiments were conducted in Python version 3.13 using HoughCircles from OpenCV version 4.11.0 [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] and gp_minimize (BO) from scikit-optimize version 0.10.2 [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003e2. Data Acquisition and Image Classes\u003c/h3\u003e\n\u003cp\u003eA total of six image classes (A\u0026ndash;F) (representative images for each class are shown in supplementary table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), each containing five images, were used to benchmark the model across a broad range of microscopy conditions. Classes A\u0026ndash;C consist of microcarriers characterised by high contrast and spherical morphology, while classes D\u0026ndash;F consist of cellular images with higher background noise and irregular shapes. Ground truth annotations were generated using the Computer Vision Annotation Tool (CVAT), where an expert manually marked the precise centre coordinates and diameters for every visible particle to establish the true population characteristics.\u003c/p\u003e\n\u003ch3\u003e3. Experimental Benchmarking\u003c/h3\u003e\n\u003cp\u003eThe performance of the trained automated system was rigorously evaluated using Leave-One-Out Cross-Validation (LOOCV) within each image class to ensure that the stable parameter configurations identified by the Bayesian framework could effectively generalise to unseen data [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. To provide a comprehensive assessment, the model was benchmarked against two standard practices that currently define the industry baseline.\u003c/p\u003e \u003cp\u003eThe first of these was a semi-automated analysis where an expert spent 30 minutes per image to manually determine the optimal binarisation and thresholding configurations using the ImageJ \u0026ldquo;Analyze Particles\u0026rdquo; plugin, version 1.54 [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. This process highlights the significant time investment and iterative nature of traditional tools which this study seeks to streamline. The second benchmark replicated common industry quality control practices through a manual expert assessment; in this case, an experienced lab technician produced a total manual count for each image and measured the diameters of 20 randomly selected particles to derive population-level estimates for the mean diameter and coefficient of variation.\u003c/p\u003e \u003cp\u003eBy comparing the automated pipeline against these manual and semi-automated workflows, we aimed to quantify the reduction in human error and the potential impact of the technology on bioprocess monitoring. All three methods were compared using a full ground-truth dataset generated via the CVAT, where an expert manually marked the precise centre coordinates and diameters of every visible particle across all six image classes.\u003c/p\u003e \u003cp\u003eThe primary metrics for comparison were the median APE in particle counts, the average diameter, and the Coefficient of Variation (CV) of the diameters. These specific metrics were selected to evaluate the system\u0026rsquo;s capacity to provide the rapid, high-throughput data required for a transition from mass-based dosing to rational surface-area-dependent process control. Detailed mathematical definitions for the intersection-over-union (IOU) thresholds used to classify detections, alongside specific error calculations, are provided in the Supplementary Information.\u003c/p\u003e"},{"header":"Results and Discussion","content":"\u003cp\u003eThe automated system\u0026rsquo;s performance was evaluated by aggregating results into two meta-groups: Microcarriers (Classes A\u0026ndash;C) and Cellular (Classes D\u0026ndash;F) (Supplementary table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The primary objective was to demonstrate a high-throughput alternative that eliminates human error and provides the geometric precision required for surface-area-based process control.\u003c/p\u003e\n\u003ch3\u003e1. Performance on Microcarrier Populations\u003c/h3\u003e\n\u003cp\u003eFor microcarrier images, the model achieved exceptional accuracy, outperforming both manual benchmarks and semi-automated ImageJ plugins across all primary metrics. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, the model recorded a median count APE of 1.2%, compared to 1.6% for manual counts and 2.2% for ImageJ. Geometric precision was similarly high, with a median diameter APE of 1.7% (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb) and a CV error of 7.8% (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe success of the system in these conditions validates the strength of the BO framework. By exploring the parameter space over 200 evaluations, the BO mechanism identified stable configurations that generalised effectively to unseen high-contrast images, positioning the system as an immediate practical replacement for an accurate and automated microcarrier particle counting. Crucially, as visualised in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ed, the system completed these analyses over 300x faster than manual methods, reducing user time from approximately ten minutes to just seconds per image, delivering a substantial high-throughput impact.\u003c/p\u003e\n\u003ch3\u003e2. Critical Boundaries and Failure Modes\u003c/h3\u003e\n\u003cp\u003eWhile the system provides a robust solution for microcarriers, performance declined in noisy cellular environments, with a median count APE of 24.7% (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). This decline defines the \"critical boundaries\" of the GHT approach. To guide future development, we identified four distinct failure modes (illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e):\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eLow Particle-Background Contrast: As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA, low-contrast images (e.g., Class D) produce few distinct boundary gradients. Since GHT detectors rely on consistent gradient vectors to cast \"votes\" for circle centres, low contrast results in insufficient voter accumulation and significant undercounting [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eNon-Particle Artefacts: Cellular images often contain smudges or debris (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). These spurious edges cause false peaks in the accumulator space, leading to false-positive detections, while simultaneously obscuring the edges of real objects [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eHeterogeneous Colour and Size: When particles exhibit diverse colours or a wide size range (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC), the BO mechanism may struggle to identify a single parameter interval that captures all instances without overgeneralising [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eHigh Degree of Clustering: Dense regions, such as the outlier image 2 in Class A, proved challenging (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD). When particles overlap, their edge points cast votes that merge into a single broad peak rather than distinct maxima. The \"minDist\" parameter, designed to prevent duplicate detections, unfortunately suppresses true detections when particles are closer than the specified threshold, leading to systematic undercounting in clustered regions [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003e3. Discussion of Impact\u003c/h3\u003e\n\u003cp\u003eThe industry\u0026rsquo;s current reliance on mass-based dosing (g/L) is a simplification that fails to account for critical variations in particle size and density, often leading to poor reproducibility and sub-optimal cell attachment. Transitioning to surface-area metrics enables a more rational process design, allowing for the systematic optimisation of seeding densities and nutrient delivery.\u003c/p\u003e \u003cp\u003eWhile manual experts can delineate indistinct boundaries, they remain limited by a ten-minute-per-sample processing time and are prone to systematic cognitive biases that introduce significant human error. This system provides an objective, reproducible baseline with a significant impact, operating 300x faster than manual methods. Although performance on complex cellular morphologies remains a challenge, the model\u0026rsquo;s 24.7% count error still significantly outperforms the 47.2% error produced by standard ImageJ thresholding. This establishes adaptive computer vision as a superior tool for bioprocess monitoring where speed and objectivity are paramount.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study successfully validated an adaptive computer vision framework that provides high precision circular particle quantification while eliminating the subjectivity and human error inherent in labour-intensive manual particle counting. By achieving a median count APE of 1.2% and performing analysis over 300x faster than experts, the system enables a critical transition toward surface-area-based process control. While complex cellular morphologies define the current \"critical boundaries\" of the GHT, this work delivers an immediate, high-throughput impact for routine microcarrier characterisation and establishes a reproducible baseline for future deep-learning-based bioprocess monitoring.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eThe authors have no conflicts of interest to declare that are relevant to the content of this article.\u003c/p\u003e\n\n\u003cp\u003eAcknowledgement\u003c/p\u003e\n\u003cp\u003eWe thank Professor Mohsen Asadnia for his support and valuable discussions. The authors acknowledge funding support from the NSW Government TechVoucher Program (Project ID: TV2500010). This project was delivered in collaboration with Macquarie University and Smarty MCs Pty Ltd.\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCox TR et al (2025) Maximising adherent cell production via customisable and dissolvable bio-polymer microcarriers. Biomed Mater, 20(5)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDing L et al (2023) Scaling up stem cell production: harnessing the potential of microfluidic devices. Biotechnol Adv 69:108271\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLevine DW, Wang DIC, Thilly WG (2004) Optimization of growth surface parameters in microcarrier cell culture. Biotechnol Bioeng 21(5):821\u0026ndash;845\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCroughan MS, Hamel JF, Wang DI (1988) Effects of microcarrier concentration in animal cell culture. Biotechnol Bioeng 32(8):975\u0026ndash;982\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGray AJ et al (2002) Cell identification and sizing using digital image analysis for estimation of cell biomass in High Rate Algal Ponds. J Appl Phycol 14(3):193\u0026ndash;204\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIgathinathane C et al (2008) Shape identification and particles size distribution from basic shape parameters using ImageJ. Comput Electron Agric 63(2):168\u0026ndash;182\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang J et al (2022) A comprehensive review of image analysis methods for microorganism counting: from classical image processing to deep learning approaches. Artif Intell Rev 55(4):2875\u0026ndash;2944\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMahony NO et al (2020) \u003cem\u003eDeep Learning vs. Traditional Computer Vision\u003c/em\u003e. Vol. 943\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Z et al (2022) A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects. IEEE Trans Neural Netw Learn Syst 33(12):6999\u0026ndash;7019\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu Z, Bajaj C (2004) Detecting circular and rectangular particles based on geometric feature detection in electron micrographs. J Struct Biol 145(1\u0026ndash;2):168\u0026ndash;180\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHassanein AS et al (2015) \u003cem\u003eA Survey on Hough Transform, Theory, Techniques and Applications.\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuen HK et al (1990) Comparative study of Hough Transform methods for circle finding. Image Vis Comput 8(1):71\u0026ndash;77\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShahriari B et al (2016) \u003cem\u003eTaking the Human Out of the Loop: A Review of Bayesian Optimization.\u003c/em\u003e Proceedings of the IEEE, 104(1): pp. 148\u0026ndash;175\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBradski G (2025) OpenCV 4.x \u0026ndash; Image Processing Module: Feature Detection (imgproc). OpenCV Developers\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003econtributors, s.-o., \u003cem\u003eskopt.gp_minimize\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDevelopers I \u003cem\u003eImageJ - Particle Analysis\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGyeongyong H, Hun C, Jihong K (2016) Robust Hough Transform with Edge Strength. Int J Eng Res and, V5(11).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"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":"bioprocess-and-biosystems-engineering","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Bioprocess and Biosystems Engineering](https://www.springer.com/journal/449)","snPcode":"449","submissionUrl":"https://submission.nature.com/new-submission/449/3","title":"Bioprocess and Biosystems Engineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Microcarrier, Bioprocessing, Adherent cell culture, Vaccine production","lastPublishedDoi":"10.21203/rs.3.rs-9391084/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9391084/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eOptimising adherent cell culture requires precise microcarrier quantification to transition from traditional mass-based (g/L) reporting to more rational surface-area-dependent process control. Current monitoring workflows are heavily hindered by human error, subjectivity, and the time-intensive nature of manual counting. The objective of this research is to provide a highly efficient, accurate and accessible technological solution that streamlines particle counting and analysis, therefore significantly contributing to the optimisation of biomanufacturing workflows. We present an adaptive, automated computer vision framework utilising the Gradient Hough Transform (GHT) integrated with a Bayesian Optimisation (BO) loop for rapid, supervised parameter tuning. For aggregated microcarrier images, the system achieves a median Absolute Percentage Error (APE) of 1.2% for counts, 1.7% for average diameter estimation, and 7.8% for population variation estimation. In these conditions, the system outperforms both manual expert benchmarks and semi-automated ImageJ plugins across every metric. While performance declines on noisy cellular images, the model's 24.7% count error still significantly outperforms the 47.2% error produced by ImageJ while remaining significantly faster than manual methods. Overall, the framework provides a rapid and reproducible alternative for microcarrier enumeration and characterisation, reducing user time from several minutes to seconds per image. While identifying \"critical boundaries\" in noisy cellular images, this work establishes a robust, user-friendly baseline for automated microcarrier characterisation in regenerative medicine and vaccine production.\u003c/p\u003e","manuscriptTitle":"A Quantitative Approach to Optimise Microcarrier Surface Area for Smarter Adherent Cell Culture","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-09 00:47:22","doi":"10.21203/rs.3.rs-9391084/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2026-04-27T02:29:42+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-16T09:56:40+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-16T09:55:16+00:00","index":"","fulltext":""},{"type":"submitted","content":"Bioprocess and Biosystems Engineering","date":"2026-04-12T01:45:28+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bioprocess-and-biosystems-engineering","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Bioprocess and Biosystems Engineering](https://www.springer.com/journal/449)","snPcode":"449","submissionUrl":"https://submission.nature.com/new-submission/449/3","title":"Bioprocess and Biosystems Engineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"3a950d09-1116-4010-9a5a-95463b3b4bf4","owner":[],"postedDate":"May 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-09T00:47:22+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-09 00:47:22","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9391084","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9391084","identity":"rs-9391084","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}