Generalizable Self-supervised Learning for Imaging Flow Cytometry on Multi-dataset Leukocyte Differential

Generalizable Self-supervised Learning for Imaging Flow Cytometry on Multi-dataset Leukocyte Differential | 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 Generalizable Self-supervised Learning for Imaging Flow Cytometry on Multi-dataset Leukocyte Differential Yi Zhang, Xukun Huang, Zirui Wang, Xinyue Du, Zimeng Fang, Songjiang Chen, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7956597/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Background Imaging flow cytometry with supervised learning can realize high-accuracy leukocyte classification. However, since supervised learning relies on annotated cell images and this labeling process generates cell losses, current imaging flow cytometry with supervised learning cannot be used for leukocyte differential. Methodology This study proposes a self-supervised contrastive learning framework pretrained on a non-annotated dataset comprising bright-field leukocyte images acquired via a custom imaging flow cytometry. The pretrained feature extractor, frozen during downstream tasks, employs lightweight linear or multilayer perceptron classifier heads to realize leukocyte differential in three independent annotated datasets: normal leukocyte subtypes (4-class), leukemia cells (2-class), and mixed normal/abnormal leukocytes (8-class). Result The proposed method achieves classification accuracies of 96.16%, 96.68%, and 92.24% across three independent datasets, performing comparably to supervised baselines (with differences of less than 2%), therefore demonstrating high accuracy for leucocyte differentials. Furthermore, since the abnormal leukocytes in the latter two datasets were not included in the pre-training dataset, the corresponding F1-scores of 96.68% and 91.92% indicate the method’s strong generalization ability. Conclusion The presented self-supervised learning framework enables high-accuracy and strong-generalization leukocyte differential using non-annotated images from imaging flow cytometry. Physical sciences/Engineering/Electrical and electronic engineering Physical sciences/Nanoscience and technology/Nanobiotechnology/Biosensors Self-supervised learning Imaging flow cytometry Leukocyte differential Strong generalization Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Leukocyte differential refers to the quantification of the percentage and absolute counts of white blood cells (WBCs) in peripheral blood, and it is widely used as a first-line hematological test [ 1 ]. The gold-standard method for leukocyte differential analysis involves the microscopic examination of blood smears, in which a blood sample is applied to slides, stained with acidic eosin and basic methylene blue dyes, and observed under a microscope [ 2 ], [ 3 ], [ 4 ]. Based on the morphological analysis of nuclear characteristics, organelle architecture, and cytoplasmic or membrane properties, trained professionals in hematology are able to classify WBCs into recognized normal types—including basophils (BAS), eosinophils (EOS), lymphocytes (LYM), monocytes (MON), and neutrophils (NEU)—as well as various abnormal forms, such as myeloblasts, promyelocytes, myelocytes, metamyelocytes, and atypical lymphocytes [ 5 ], [ 6 ], [ 7 ]. Although widely used in clinical practice, this method is time-consuming, which seriously limits diagnostic efficiency [ 8 ], [ 9 ]. To address this issue, hematology analyzers were designed to perform continuous analysis of single blood cells based on light scattering and impedance variations, enabling high-throughput leukocytes differentiation [ 10 ], [ 11 ], [ 12 ]. Automated commercial hematology analyzers, such as the Sysmex XN-2000 and Beckman Coulter DxH 900, are routinely utilized in clinical laboratories worldwide [ 13 ], [ 14 ], [ 15 ], [ 16 ]. Unfortunately, accurate leukocyte differential was achieved only for normal WBCs, but not for abnormal ones, due to the absence of morphological parameters [ 17 ]. To overcome these limitations, imaging flow cytometry was utilized to capture fluorescence and/or bright-field images of white blood cells for classification by combining machine learning and deep learning [ 18 ], [ 19 ], [ 20 ], [ 21 ]. For example, the commercial ImageStream platform attained 97.0% F1-scores in four-part leukocyte classification with machine learning [ 22 ]. In 2024, Wang, et al. [ 23 ] accomplished a classification accuracy of 97.1% for three ovarian cell types by integrating three supervised networks on brightfield and three-channel fluorescence imaging data. In 2025, Chen, et al. [ 24 ] realized a classification accuracy of 99.6% for four leukocyte subtypes through five deep neural network fusion strategies of supervised LSTM and ResNet-18 applied to impedance and fluorescence imaging data. However, since supervised learning relies on annotated cell images and this labeling process generates cell losses, current imaging flow cytometry with supervised learning cannot be used for leukocyte differential. Thus, label-free bright-field images captured by imaging flow cytometry [ 25 ], [ 26 ], [ 27 ] coupled with self-supervised learning [ 28 ], [ 29 ], [ 30 ] holds a significant promise for leukocyte differential. Different from supervised learning, self-supervised learning can utilize unlabeled data for training and classification by automatically generating proxy tasks to generate supervisory signals from the data itself, thereby obtaining guidance from large-scale data and achieving highly discriminative features for downstream classification [ 31 ], [ 32 ]. Within the field of leukocyte classification of blood smears, self-supervised learning has demonstrated high accuracies and strong generalization. For example, in 2024, Jia, et al. [ 33 ] achieved a classification accuracy of 93.51% across 15 classes on the TMAMD leukocyte-stained blood smear dataset using a masked autoencoder pretrained on a stained blood smear dataset. In 2024, Koch, et al. [ 34 ] reported a classification accuracy of 86.6% for 10-part leukocyte differential on the Acevedo dataset, utilizing a pretrained DINO v2 network. As to leukocyte classification based on imaging flow cytometry and self-supervised learning, the only previous report was in 2023, in which Zhang, et al. [ 35 ] reported an accuracy of 91.9% for three-part classification of normal leukocytes using a pretrained deep convolutional autoencoder with frozen parameters. However, the generative learning used in this study focused on pixel-level reconstruction, which over-emphasized low-level details irrelevant to semantic discrimination, thereby potentially compromising the accuracy and generalizability of leukocyte classification. This study establishes a self-supervised contrastive framework for leukocyte differential, based on bright-field images from imaging flow cytometry. Compared to generative learning, contrastive learning directly enhances feature separability and robustness against perturbations without being constrained by the need for precise input reconstruction. More specifically, the feature extraction network was pretrained on unannotated leukocyte images and then used for three downstream leukocyte differential tasks, including those involving previously unseen subtypes with lightweight linear or MLP classifiers, while freezing the feature extractor parameters. The MoCo v2-based methodology achieves performance parity with supervised baselines, underscoring exceptional cross-dataset generalization capabilities without architectural retraining. The unannotated and strongly generalizable imaging flow cytometry-based sorting method may enhance the capability of clinical leukocyte differential. 2. Materials and methodology 2.1 Working flow chart The overall workflow of this study, as illustrated in Fig. 1 , comprised three major phases: (a) image acquisition, (b) self-supervised pre-training, and (c) muti-dataset downstream leukocyte differential. Initially, leukocyte samples—including mixed leukocytes, purified leukocytes, and leukemia cell lines—were captured using a home-developed imaging flow cytometry system. Through multiple independent experiments, one non-annotated dataset (derived from leukocyte mixtures) and three annotated datasets were constructed for pre-training and generalization evaluation. In the self-supervised pre-training phase, a Momentum Contrast (MoCo v2) framework [ 36 ] was employed to learn representative features from the non-annotated dataset. This framework incorporated a ResNet-18 [ 37 ] architecture as the shared backbone network, consisting of a query encoder (updated via backpropagation) and a momentum encoder (updated via momentum-based moving average). The pre-trained query encoder was subsequently saved as a fixed feature extractor. For the downstream differential phase, the pre-trained weights were frozen, and only a lightweight linear classifier or multilayer perceptron (MLP) head was trained and evaluated on each of the three annotated datasets to validate cross-dataset generalization. 2.2 Image acquisition and dataset composition The image data in this study were acquired by a home-developed imaging flow cytometry (see Fig. 1 (a)). The cytometry utilized a PZT-based acoustic focusing method, which ensured that cells pass sequentially through the depth-of-field region at the center of the channel, enabling high-throughput acquisition of high-quality, in-focus bright-field cell images. The samples tested included leukocyte mixtures, purified leukocytes, and leukemia cell lines. For detailed configurations regarding materials, cell preparation, and the imaging flow cytometry setup, please refer to our group’s previous publication [ 38 ]. Through multiple experiments, four mutually independent datasets were constructed: one non-annotated dataset for self-supervised pre-training and three annotated datasets for downstream differential tasks to test generalization of pre-training feature exactor. No cell images were reused across these datasets. The self-supervised pre-training dataset consisted of images from leukocyte mixtures obtained from four healthy volunteers and contained only non-annotated images of normal leukocytes. The three downstream differential datasets all comprised annotated leukocyte images. Normal leukocyte types came from purified leukocytes isolated via fluorescence-activated cell sorting after red blood cell lysis of the whole blood, while abnormal leukocyte types were derived from in vitro cultured leukemia cell lines. All three datasets are independent and non-overlapping. The first dataset included the four most abundant normal leukocyte types—EOS (eosinophil), LYM (lymphocyte), MON (monocyte), and NEU (neutrophil)—and served as a four-class baseline for normal cell classification. The second dataset contained two types of leukemia cell lines, HL60 (promyelocytic leukemia) and K562 (chronic myelogenous leukemia), which were unseen during feature extractor pre-training. It was used to validate the generalizability of the feature exactor in classifying novel abnormal cell types. The third dataset comprised five types of normal leukocytes—BAS (basophil), EOS, LYM, MON, and NEU—as well as three types of abnormal leukemia cell lines: Jurkat (T-cell leukemia), HL60, and K562. This dataset included both all categories of normal cells and abnormal cell types (representing both lymphoid and myeloid lineages) not encountered during feature extractor pre-training. 2.3 Self-supervised Pre-training The self-supervised learning workflow in this study consisted of two parts. First, a feature extractor was pre-trained using self-supervised learning on a large amount of non-annotated data. Then, for the three downstream differential tasks, the deep neural network for feature extraction was frozen, and only a simple classifier based on the extracted features was trained separately using the annotated datasets. This approach completed cell classification with minimal training cost while keeping the backbone feature extraction deep neural network unchanged. The training in this study was performed on Ubuntu 18.04 systems using PyTorch 1.8.2 and CUDA 11.1, hosted on AutoDL servers with an Intel Xeon Gold 6430 CPU (16 vCPUs) and a single NVIDIA RTX 4090 GPU (24 GB). In the self-supervised pre-training phase, as illustrated in Fig. 1 (b), this study employed the MoCo v2 architecture for self-supervised contrastive learning. Its key features included a momentum encoder and a queue dictionary for maintaining a large set of negative samples. The backbone network for feature extraction was the convolutional neural network ResNet-18, which offers a good balance between model complexity and accuracy, effectively meeting the feature extraction requirements for leukocyte images. Since the effectiveness of contrastive learning depended on the data augmentation strategy, this study designed a composite augmentation scheme including rotation, flipping, erosion, and Gaussian blur, tailored to the characteristics of non-annotated bright-field leukocyte images, which were single-channel and sensitive to magnification. The batch size was set to 256, the maximum feasible on the RTX 4090 GPU (24 GB), to ensure high training speed. Key hyperparameters, including learning rates of 0.01 and 0.05, then intermediate values of 0.02 and 0.025, and finally 0.0225, were tuned to identify the optimal network for downstream feature extraction. The training performance of MoCo v2 was evaluated using two primary metrics: Top-1 accuracy, where higher values indicated better representation quality upon convergence, and contrastive loss, where lower values signified more effective learning of invariant features during pre-training until convergence was achieved. 2.4 Multi-dataset Downstream Leukocyte Differential In the three downstream leukocyte differential tasks, as shown in Fig. 1 (c), based on the feature extractor pre-trained with the aforementioned MoCo v2 self-supervised approach, two schemes were employed for training (with 5-fold cross-validation): a single linear classification head and a 3-layer MLP. In the linear head scheme, the feature extractor was frozen, and only a single layer (the linear head) was trained. Since no new features were introduced into the feature space, only linear division of existing features occurred, its classification results could demonstrate the effectiveness of the features extracted by self-supervised learning. Introducing a 3-layer MLP, which froze the feature extraction network and used only a simple non-linear classification network for fine-tuning, could improve the classification accuracy of downstream tasks based on features extracted by self-supervised learning with minimal training cost compared to retraining the whole deep neural networks. To contrast the effectiveness of the above scheme, this study also employed another self-supervised learning scheme and three supervised learning schemes for comparison on the three downstream tasks. All tasks used 5-fold cross-validation, and key hyperparameters like learning rate were tuned for optimal results. A batch size of 256 was used for the fastest training speed on the available GPU. SimCLR was used as another self-supervised scheme. Its effectiveness was measured by freezing the feature-extracting ResNet-18 and training the same linear classification head on the three downstream tasks, comparing it to the feature extraction network trained with MoCo v2 to demonstrate differences in feature effectiveness among self-supervised schemes. The SimCLR self-supervised feature extractor training also used ResNet-18 as the backbone network with the same data augmentation as MoCo v2. However, as SimCLR lacks momentum update and its effectiveness relies on a large batch size, and the computational power of the single RTX 4090 (24 GB) was limited, gradient accumulation was used. The original batch size was 256 (the maximum on the RTX 4090), with a gradient accumulation of 16, resulting in an effective batch size of 4096 to meet SimCLR's training requirements. Other key hyperparameters included a temperature of 0.5 and a learning rate of 0.1. For the downstream linear classification head training, the ResNet-18 backbone network trained by SimCLR was frozen, and only the single neural network layer was trained, exactly the same as the linear classification head used with the MoCo v2-trained feature extractor. The supervised learning methods compared on the three downstream tasks were trained separately using the annotated datasets from the respective downstream task. The employed supervised learning methods included MLP, ResNet-18, and ViT-small. MLP is a multilayer perceptron, a feedforward artificial neural network model composed of an input layer, hidden layers, and an output layer, capable of learning complex patterns and relationships in data. Using only MLP for supervised learning allowed comparison with using the self-supervised feature extractor followed by MLP, thus demonstrating the effectiveness of the self-supervised learned features. ResNet-18, as a representative convolutional neural network with balanced performance, is an 18-layer deep residual network that effectively alleviates the vanishing gradient problem in deep networks through skip connections. It was a lightweight convolutional neural network that achieves high accuracy with a relatively small number of model parameters and served as a comparison for the supervised learning convolutional neural network architecture against the scheme in this study. Vision Transformer (ViT), proposed by Google, is a Transformer-based network for computer vision. It divides images into fixed-size patches and treats them as token sequences, extracting features through a Transformer encoder to handle visual tasks. ViT-small, as the smallest model, had a number of network parameters roughly on par with ResNet-18 and represented the supervised learning ViT architecture compared to the scheme in this study. In the comparison of results from the three downstream tasks, classification accuracy, precision, recall, and F1-score were calculated for evaluation. Accuracy measured the proportion of correctly predicted samples. Precision focused on the proportion of true positive samples among those predicted as positive. Recall focused on the proportion of correctly predicted samples among all true positive samples. F1-score was the harmonic mean of precision and recall, providing a comprehensive assessment of model performance. For the core scheme in this study, the self-supervised scheme based on MoCo v2, its results on the three downstream tasks using both the linear classification head and the 3-layer MLP classification head were additionally evaluated using confusion matrices. The confusion matrix is a tool that visualizes the correspondence between predicted results and true samples in a tabular format, used for detailed analysis of the types of classification errors. Each cell within the matrix contains two values: the absolute count of instances and a corresponding proportion. The proportion is calculated by taking the number of cases in a given cell and dividing it by the total number of instances associated with the true class of that row, thereby providing a normalized view of the prediction distribution for each actual class. 3. Results and discussion 3.1 Datasets The details of the image datasets captured by the home-developed imaging flow cytometry were shown in Table 1 . It started with the leukocyte mixture images used for the self-supervised feature extraction task. These were sourced from 4 healthy donors, with a total of 80,347 images. This portion of cell images was obtained directly from the leukocyte mixture after the lysis of red blood cells in whole blood. They were non-annotated, used for training a stable and effective feature extractor in the self-supervised task. Representative images were shown in Fig. 2 (a). Table 1 Leukocyte datasets captured by a home-developed imaging flow cytometry. Dataset Name Primary Use Annotate Status Total Cell Classes & Quantity Leukocyte Mixture Dataset Self-supervised Pre-training Non-annotated 80,347 Normal Leukocytes: 80,347 4-class Normal Leukocyte Dataset Task 1: 4-class Normal Leukocyte Differential Annotated 24,834 1. Eosinophil (EOS): 4,175 2. Lymphocyte (LYM): 5,879 3. Monocyte (MON): 6,383 4. Neutrophil (NEU): 8,397 2-class Abnormal Leukocyte Dataset Task 2: 2-class Abnormal Leukocyte Differential Annotated 11,093 1. HL60: 5,468 2. K562: 5,625 8-class Normal/Abnormal Leukocyte Dataset Task 3: 8-class Normal/Abnormal Leukocyte Differential Annotated 8,419 1. Basophil (BAS): 515 2. Eosinophil (EOS): 1,034 3. Lymphocyte (LYM): 1,159 4. Monocyte (MON): 1,189 5. Neutrophil (NEU): 1,238 6. Jurkat: 1,086 7. HL60: 1,061 8. K562: 1,137 Then, annotated leukocyte image data were captured and subsequently used in the three downstream differential tasks. The normal leukocyte types were derived from purified leukocytes, which were obtained via fluorescence-activated cell sorting (FACS) after red blood cell lysis in whole blood from healthy donors. The abnormal leukocyte types were derived from in vitro cultured leukemia cell lines. The three datasets were based on mutually independent data with no overlap. Representative images were shown in Fig. 2 (b) – (d). Dataset 1 (Fig. 2 (b)) contained 4 types of purified normal leukocytes: EOS (eosinophils), LYM (lymphocytes), MON (monocytes), and NEU (neutrophils), all from human peripheral blood of healthy individuals. The total data volume was relatively large at 24,834. Dataset 2 (Fig. 2 (c)) contained 2 types of leukemia cells: HL60 and K562, both of which were abnormal cell types not present during the training of the self-supervised learning feature extractor. The total data volume was moderate at 11,093. Dataset 3 (Fig. 2 (d)) contained 8 types of normal and abnormal leukocytes: 5 types of purified normal leukocytes (BAS (basophils), EOS (eosinophils), LYM (lymphocytes), MON (monocytes), NEU (neutrophils)) and 3 types of leukemia cells (HL60, Jurkat, and K562). These three leukemia cell types were also not present during the training of the self-supervised learning feature extractor. The total data volume was the smallest at 8,419, but it had the most categories. Except for BAS (basophils), which were fewer at 515, the other categories had around 1,000 images each to ensure relatively balanced data volume. Judging from the example images, they were all clear leukocyte bright-field images without defocusing, and the cell contours and internal textures were distinct. Based on the visually apparent cell size, normal leukocytes were smaller than abnormal leukemia cells, while the size of normal leukocytes remained consistent across datasets. Among the normal leukocytes, lymphocytes were significantly smaller than the other four types, which was consistent with standard leukocyte analysis. The leukocyte mixture contained cells with both larger and smaller diameters, suggesting that it included all categories of normal leukocytes. 3.2 Results of Self-supervised Pre-training Figure 3 presented the self-supervised pre-training results on the non-annotated leukocyte mixture dataset, specifically comparing the performance of MoCo v2 under different learning rates. This included the variation of Top-1 accuracy with training epochs (Fig. 3 (a)) and the variation of contrastive loss with training epochs (Fig. 3 (b)). The learning rates tested were 0.01 (red), 0.02 (blue), 0.0225 (green), 0.025 (purple), and 0.05 (yellow). These results indicated that the model converged around 100 epochs under all five learning rates, justifying the selection of 200 epochs as the training duration. Among the five learning rates, the network selected for feature extraction in downstream tasks was chosen based on the highest Top-1 accuracy. The final selection was the model trained with a learning rate of 0.0225 (green curve), as it achieved the best Top-1 accuracy (99.99%), outperforming the 99.97% achieved with 0.01, 99.98% with 0.02, 99.98% with 0.025, and the 99.94% achieved with 0.05. Regarding the contrastive loss, since this study used a self-constructed dataset, the final converged contrastive loss value of approximately 6.2 only indicated effective convergence of self-supervised learning on this specific dataset and was not directly comparable to contrastive loss values reported on other datasets. Note that the final performance in terms of both Top-1 accuracy and contrastive loss were highly similar across the learning rates, achieving an extremely high final accuracy (close to 100%) and a contrastive loss value approaching 6.2. This demonstrated the parameter robustness of MoCo v2 on this dataset and confirmed the high compatibility of the selected data augmentation strategy, the ResNet-18 backbone network, and the leukocyte mixture dataset, marking a successful phase in self-supervised pre-training. 3.3 Results of Multi-dataset Downstream Leukocyte Differential After obtaining a feature extraction network through self-supervised pre-training, the network was frozen and applied to differential tasks on three downstream leukocyte datasets. For each dataset, a linear classification head was first trained to validate the effectiveness of the features, followed by training an MLP classification head to improve classification accuracy. The classification confusion matrix results were shown in Fig. 4 . In the 4-class normal leukocyte differential task (see Fig. 4 (a)), the linear classification head achieved high precision with each cell type recall exceeding 0.9, demonstrating the effectiveness of feature extraction. The MLP classification head further improved the recall for each cell type to above 0.94 (see Fig. 4 (b)). In terms of cell types, the two lowest recall values were for EOS and NEU, which aligns with the expectation that eosinophils and neutrophils are similar and thus difficult to distinguish. In the 2-class abnormal leukemia cell differential task (Fig. 4 (c) - (d)), the model exhibited exceptional accuracy in distinguishing between HL60 and K562, two types of abnormal leukocytes not present in the pre-training dataset, with each class recall exceeding 0.95, effectively proving the generalization capability of the self-supervised pre-trained feature extraction network. The MLP classification head effectively reduced the minor misclassifications observed with the linear classification head, achieving a class-wise recall exceeding 0.97. The high accuracy might be attributed to the relative simplicity of the binary classification task. In the most complex 8-class normal/abnormal mixed leukocyte differential task (Fig. 4 (e) – (f)), the linear classification head still achieved a minimum recall of 0.84 for NEU, more robustly demonstrating the generalization of the feature extraction network. The recall for each cell type was relatively lower due to the task's complexity. In the confusion matrix for linear classification shown in Fig. 4 (e), LYM exhibited the highest recall of 0.96, likely because their significantly smaller diameter makes them a distinctive feature. They were followed by three types of abnormal leukocytes, which had larger diameters than normal leukocytes. The remaining four types of normal leukocytes, with similar diameters, showed lower recall. These results were consistent with cell diameter being the most obvious and physically meaningful distinguishing characteristic. The total number of misclassified samples between normal and abnormal cells was 146, which was less than the total number of misclassified samples within normal leukocytes (573), verifying that greater feature differences exist between normal and abnormal cells. As shown in Fig. 4 (f), the MLP classification head significantly improved the recall for several leukocyte types (e.g., 0.90 of MON, 0.87 of NEU, and 0.96 of Jurkat). However, the improvement for BAS was limited, with only two additional correctly classified samples, likely because the number of BAS samples was only about half that of other leukocytes and the MLP classification head had a simple structure with limited capacity. Despite this, high-accuracy differential with each leukocyte type recall exceeding 85% was achieved. Table 2 presented the performance of the proposed self-supervised learning scheme compared to other different supervised and self-supervised methods on three leukocyte datasets, evaluating Accuracy, Precision, Recall, and F1-score to comprehensively showcase model performance. First, the supervised learning baseline, which involved direct MLP classification on the data, performed significantly lower than other methods, with accuracies of only 43.40% and 43.37% for the 4-class Normal and 8-class Normal/Abnormal tasks, respectively. This demonstrated that simply using an MLP or even simpler linear layers could not effectively distinguish leukocytes, thus highlighting the necessity and effectiveness of feature extraction through self-supervised pre-training when compared to training classification heads afterwards. Table 2 Classification performance of different supervised/self-supervised methods on three leukocyte datasets. Dataset Methodology Accuracy Precision Recall F1-score 4-class Normal MLP (Sup) 43.40% 45.33% 36.04% 33.00% ResNet-18 (Sup) 96.81% 96.80% 96.79% 96.78% ViT-Small (Sup) 97.61% 97.63% 97.47% 97.54% SimCLR (Linear Head) 73.58% 72.96% 72.68% 72.81% MoCo v2 (Linear Head) 93.56% 93.70% 93.45% 93.57% MoCo v2 (MLP Head) 96.16% 96.15% 95.98% 96.06% 2-class Abnormal MLP (Sup) 85.28% 85.29% 85.26% 85.27% ResNet-18 (Sup) 97.67% 97.68% 97.67% 97.67% ViT-Small (Sup) 95.17% 95.17% 95.16% 95.17% SimCLR (Linear Head) 95.46% 95.45% 95.46% 95.46% MoCo v2 (Linear Head) 95.47% 95.46% 95.47% 95.46% MoCo v2 (MLP Head) 96.68% 96.69% 96.68% 96.68% 8-class Normal/ Abnormal MLP (Sup) 43.37% 43.13% 40.34% 39.63% ResNet-18 (Sup) 94.01% 93.83% 94.11% 93.88% ViT-Small (Sup) 92.54% 92.49% 92.25% 92.37% SimCLR (Linear Head) 72.70% 71.87% 71.40% 71.46% MoCo v2 (Linear Head) 91.46% 91.12% 91.15% 91.13% MoCo v2 (MLP Head) 92.24% 92.00% 91.87% 91.92% Subsequently, two representative supervised learning networks, ResNet-18 and ViT-Small, provided performance benchmarks on the three datasets. It was important to note that supervised learning required retraining the entire feature extraction network separately on each annotated dataset and could not leverage non-annotated data. Then, two self-supervised learning contrastive methods, SimCLR and MoCo v2, were evaluated based on linear classification head performance. MoCo v2 outperformed SimCLR by approximately 20% across all four metrics (Accuracy, Precision, Recall, F1-score) on both the 4-class Normal and 8-class Normal/Abnormal tasks, proving that the features learned by MoCo v2 through self-supervised learning possess higher discriminability and generalization. Finally, by freezing the feature extraction network pre-trained with MoCo v2 and training only a simple MLP classification head, the achieved performance on all four metrics was within approximately 2% of the supervised learning benchmarks across all three downstream differential tasks. This fully demonstrated that the self-supervised pre-trained feature extraction network possesses exceptional cross-dataset generalization capability. 4. Conclusion This study establishes a self-supervised contrastive framework for leukocyte differential, based on bright-field images from imaging flow cytometry. By pretraining on non-annotated data, the feature extractor learned discriminative representations invariant to augmentation, achieving high performance on downstream tasks with minimal fine-tuning. The method generalizes across diverse datasets, including unseen abnormal leukocyte types, with accuracy rivaling supervised models. This underscores the potential of self-supervised learning to enhance flow cytometry by reducing manual annotation burdens and improving model generalization capability. Future work will extend to broader representative patient samples, classification of rare cell populations, and multimodal diagnostic integration including fluorescence labeling or impedance signals, promoting the clinical application of intelligent diagnosis in flow cytometry. Declarations Competing interests The authors declare no conflicts of interest. Acknowledgement This work was supported from the National Natural Science Foundation of China (Grants No.: 62331025 and 62121003) and the Beijing Municipal Natural Science Foundation (Grants No.: 4232010). References J. A. Molnar, "12 - Automated blood cell analysis," Rodak's Hematology (Sixth Edition) , E. M. Keohane, C. N. Otto and J. M. Walenga, eds., pp. 174–200, St. Louis (MO): Elsevier, 2020. J. E. Lewis, and O. Pozdnyakova, “Digital assessment of peripheral blood and bone marrow aspirate smears,” International Journal of Laboratory Hematology , vol. 45, pp. 50–58, Jun. 2023. M. L. Chase et al., “Consensus recommendations on peripheral blood smear review: Defining curricular standards and fellow competency,” Blood Advances , vol. 7, no. 13, pp. 3244–3252, Jul. 2023. M. O. Nakashima, S. N. Coulter, B. J. Blond, R. W. Brown, and J. A. Vos, “The impact of pathologist review on peripheral blood smears a college of american pathologists q-probes study of 22 laboratories,” Archives of Pathology & Laboratory Medicine , vol. 149, no. 7, pp. 618–626, Jul. 2025. L. Palmer et al., “ICSH recommendations for the standardization of nomenclature and grading of peripheral blood cell morphological features,” International Journal of Laboratory Hematology , vol. 37, no. 3, pp. 287–303, Jun. 2015. C. Matek, S. Schwarz, K. Spiekermann, and C. Marr, “Human-level recognition of blast cells in acute myeloid leukaemia with convolutional neural networks,” Nature Machine Intelligence , vol. 1, no. 11, pp. 538–544, Nov. 2019. C. Vilchez, "13 - Examination of the peripheral blood film and correlation with the complete blood count," Rodak's Hematology (Sixth Edition) , E. M. Keohane, C. N. Otto and J. M. Walenga, eds., pp. 201–218, St. Louis (MO): Elsevier, 2020. Z. X. Tian, Y. C. Wei, Y. L. Yu, F. L. Zhou, and Z. L. Huang, “Blood cell analysis: From traditional methods to super-resolution microscopy,” Photonics , vol. 9, no. 4, pp. 261, Apr. 2022. W. X. J. Cheng et al., “Application of image recognition technology in pathological diagnosis of blood smears,” Clinical and Experimental Medicine , vol. 24, no. 1, pp. 181, Aug. 2024. S. Rastogi et al., “Overall equipment effectiveness, efficiency and slide review analysis of high-end hematology analyzers,” Practical Laboratory Medicine , vol. 30, pp. e00275, May. 2022. M. Daves et al., “Modern hematology analyzers: Beyond the simple blood cells count (with focus on the red blood cells),” Journal of Laboratory and Precision Medicine , vol. 9, pp. 4, Jan. 2024. L. A. Liu et al., “Multi-wavelength multi-direction laser light scattering for cell characterization using machine learning-based methods,” Cytometry Part A , vol. 103, no. 10, pp. 796–806, Oct. 2023. S. Buoro et al., “Multicenter evaluation of blood cell counts on 8 automated hematology analyzers,” Biochimica Clinica , vol. 40, no. 3, pp. 195–203, Sep. 2016. O. Ciepiela et al., “A comparison of Mindray BC-6800, Sysmex XN-2000, and Beckman Coulter LH750 automated hematology analyzers: A pediatric study,” Journal of Clinical Laboratory Analysis , vol. 30, no. 6, pp. 1128–1134, Nov. 2016. L. Meintker, J. Ringwald, M. Rauh, and S. W. Krause, “Comparison of automated differential blood cell counts from Abbott Sapphire, Siemens Advia 120, Beckman Coulter DxH 800, and Sysmex XE-2100 in normal and pathologic samples,” American Journal of Clinical Pathology , vol. 139, no. 5, pp. 641–650, May. 2013. E. Schapkaitz, and S. Raburabu, “Performance evaluation of the new measurement channels on the automated Sysmex XN-9000 hematology analyzer,” Clinical Biochemistry , vol. 53, pp. 132–138, Mar. 2018. M. R. Wang et al., “Developments of conventional and microfluidic flow cytometry enabling high-throughput characterization of single cells,” Biosensors-Basel , vol. 12, no. 7, pp. 443, Jul. 2022. Q. Huang et al., “Imaging flow cytometry: From high - resolution morphological imaging to innovation in high - throughput multidimensional biomedical analysis,” Frontiers in Bioengineering and Biotechnology , vol. 13, pp. 12, May. 2025. V. Railean, and B. Buszewski, “Flow cytometry - sophisticated tool for basic research or/and routine diagnosis; impact of the complementarity in both pre- as well as clinical studies,” Critical Reviews in Analytical Chemistry , vol. 54, no. 7, pp. 2087–2109, Oct. 2024. D. Vonficht et al., “Ultra-high-scale cytometry-based cellular interaction mapping,” Nature Methods , vol. 22, no. 9, Aug. 2025. H. Kanno et al., “High-throughput fluorescence lifetime imaging flow cytometry,” Nature Communications , vol. 15, no. 1, pp. 7376, Sep. 2024. M. Nassar et al., “Label-free identification of white blood cells using machine learning,” Cytometry Part A , vol. 95, no. 8, pp. 836–842, May. 2019. Z. W. Wang, Q. Liu, J. Zhou, and X. T. Su, “Single-detector multiplex imaging flow cytometry for cancer cell classification with deep learning,” Cytometry Part A , vol. 105, no. 9, pp. 666–676, Sep. 2024. S. Chen et al., “Bimodal fusion of single-cell impedance and imaging data using five deep neural networks,” IEEE Nanotechnology Magazine , vol. 19, no. 1, pp. 24–30, Jan. 2025. Z. Q. Du et al., “Label-free detection and enumeration of rare circulating tumor cells by bright-field image cytometry and multi-frame image correlation analysis,” Lab on a Chip , vol. 22, no. 18, pp. 3390–3401, Sep. 2022. C. F. Otesteanu et al., “A weakly supervised deep learning approach for label-free imaging flow-cytometry-based blood diagnostics,” Cell Reports Methods , vol. 1, no. 6, pp. 100094, Oct. 2021. Z. W. Wang, Q. Liu, R. Chu, K. Song, and X. T. Su, “Single-detector dual-modality imaging flow cytometry for label-free cell analysis with machine learning,” Optics and Lasers in Engineering , vol. 168, pp. 107665, Sep. 2023. L. L. Zeune et al., “Deep learning of circulating tumour cells,” Nature Machine Intelligence , vol. 2, no. 2, pp. 124–133, Feb. 2020. B. Liu et al., “Self-supervised learning reveals clinically relevant histomorphological patterns for therapeutic strategies in colon cancer,” Nature Communications , vol. 16, no. 1, pp. 2328, Mar. 2025. T. Hong, M. Peng, Y. Kim, H. E. Schellhorn, and Q. Fang, “Automated cell profiling in imaging flow cytometry with annotation-efficient learning,” Optics & Laser Technology , vol. 181, pp. 111992. 2025. Y. N. Lu, L. J. Wen, J. Z. Liu, Y. J. Liu, and X. M. Tian, “Self-supervision can be a good few-shot learner,” in Proc. Computer Vision (ECCV 2022) , Tel Aviv, Israel, Nov. 2022, pp. 740–758. V. Rani, S. T. Nabi, M. Kumar, A. Mittal, and K. Kumar, “Self-supervised Learning: A Succinct Review,” Archives of Computational Methods in Engineering , vol. 30, no. 4, pp. 2761–2775, May. 2023. N. Jia et al., “A fine-grained image classification algorithm based on self-supervised learning and multi-feature fusion of blood cells,” Scientific Reports , vol. 14, no. 1, pp. 22964, Oct. 2024. V. Koch et al., “DinoBloom: A Foundation Model for Generalizable Cell Embeddings in Hematology,” in Proc. Medical Image Computing and Computer Assisted Intervention (MICCAI 2024) , Marrakesh, Morocco. 2024, pp. 520–530. Z. Zhang et al., “Interpretable unsupervised learning enables accurate clustering with high-throughput imaging flow cytometry,” Scientific Reports , vol. 13, no. 1, pp. 20533, Nov. 2023. K. M. He et al., “Momentum contrast for unsupervised visual representation learning,” in Proc. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020) , Seattle, WA, USA, Jun. 2020, pp. 9726–9735. K. M. He, X. Y. Zhang, S. Q. Ren, J. Sun, and Ieee, “Deep residual learning for image recognition,” in Proc. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , Las Vegas, NV, USA, Jun. 2016, pp. 770–778. X. Huang et al., “Label-free 8-part leukocyte differential enabled by imaging flow cytometry of acoustic focusing and deep neural network of ResNet-18,” Biomedical Instrumentation , vol. 1, no. 1, pp. 10004, Aug. 2025. Additional Declarations There is no conflict of interest Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: revise 29 Dec, 2025 Review # 3 received at journal 25 Dec, 2025 Review # 1 received at journal 16 Dec, 2025 Reviewer # 3 agreed at journal 06 Dec, 2025 Reviewer # 2 agreed at journal 02 Dec, 2025 Reviewer # 1 agreed at journal 30 Nov, 2025 Reviewers invited by journal 07 Nov, 2025 Submission checks completed at journal 28 Oct, 2025 Editor assigned by journal 25 Oct, 2025 First submitted to journal 25 Oct, 2025 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-7956597","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":541403507,"identity":"36797b24-366b-4621-ae62-4927110af7a9","order_by":0,"name":"Yi Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8klEQVRIiWNgGAWjYDACCQjJI8/ewHCAJC1yhj0HSNPCYMxwI4FId8nPbn728GubRWLjzMcPD1fU2NkzsB9+wPBzB24tjHOOmRvLtkkktkunGRw8cyw5sYEnzYCx9wxuLcwSCWbSkkAtjbMTDA42NjADnZfDwMzYhlsLm0T6N7CWhpvHPwC11Nsz8L/Br4VHIsdM8mObBND7PCBbDjM2SBCwRUIip0ya4RwokHMKDjYcO57YJvHM4GAvHi3yM9K3Sf4oqwNG5fHNHxtqqu35+ZMfPviJRws4CHjZkH0HxAfwawAG9I8/hJSMglEwCkbBiAYASN5OYD5itwYAAAAASUVORK5CYII=","orcid":"https://orcid.org/0009-0009-9196-6508","institution":"State Key Laboratory of Transducer Technology, Aerospace Information Research Institute, Chinese Academy of Sciences","correspondingAuthor":true,"prefix":"","firstName":"Yi","middleName":"","lastName":"Zhang","suffix":""},{"id":541403508,"identity":"05a1c334-bc8a-49dd-bf86-125d98420fff","order_by":1,"name":"Xukun Huang","email":"","orcid":"","institution":"State Key Laboratory of Transducer Technology, Aerospace Information Research Institute, Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Xukun","middleName":"","lastName":"Huang","suffix":""},{"id":541403509,"identity":"a5e8a6e9-9d38-4548-89d2-7b85283e6014","order_by":2,"name":"Zirui Wang","email":"","orcid":"","institution":"Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China","correspondingAuthor":false,"prefix":"","firstName":"Zirui","middleName":"","lastName":"Wang","suffix":""},{"id":541403510,"identity":"480845cf-4621-4bb7-83cb-51353b3e031e","order_by":3,"name":"Xinyue Du","email":"","orcid":"","institution":"State Key Laboratory of Transducer Technology, Aerospace Information Research Institute, Chinese Academy of Sciences. School of Future Technology, University of Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Xinyue","middleName":"","lastName":"Du","suffix":""},{"id":541403511,"identity":"db664dd4-ba15-43fc-9f2a-df750f585637","order_by":4,"name":"Zimeng Fang","email":"","orcid":"","institution":"School of Electronic Engineering, Beijing University of Posts and Telecommunications","correspondingAuthor":false,"prefix":"","firstName":"Zimeng","middleName":"","lastName":"Fang","suffix":""},{"id":541403512,"identity":"28d27732-434a-4bcb-9df2-b9b60d6708ed","order_by":5,"name":"Songjiang Chen","email":"","orcid":"","institution":"State Key Laboratory of Transducer Technology, Aerospace Information Research Institute, Chinese Academy of Sciences.School of Electronic, Electrical and Communication Engineering, University of Chi","correspondingAuthor":false,"prefix":"","firstName":"Songjiang","middleName":"","lastName":"Chen","suffix":""},{"id":541403513,"identity":"3fac2d7c-5700-4aee-8ff0-f43620ebc1de","order_by":6,"name":"Yongfan Chen","email":"","orcid":"","institution":"State Key Laboratory of Transducer Technology, Aerospace Information Research Institute, Chinese Academy of Sciences.School of Electronic, Electrical and Communication Engineering, University of Chi","correspondingAuthor":false,"prefix":"","firstName":"Yongfan","middleName":"","lastName":"Chen","suffix":""},{"id":541403514,"identity":"a0ae5d30-13f5-42d2-839d-895d31a8854c","order_by":7,"name":"Junbo Wang","email":"","orcid":"","institution":"State Key Laboratory of Transducer Technology, Aerospace Information Research Institute (AIR), Beijing, China School of Electronic, Electrical, and Communication Engineering, University of Chinese A","correspondingAuthor":false,"prefix":"","firstName":"Junbo","middleName":"","lastName":"Wang","suffix":""},{"id":541403515,"identity":"8bfac98f-a01c-4ed2-a86f-8c54af5400d1","order_by":8,"name":"Xiaoye Huo","email":"","orcid":"","institution":"State Key Laboratory of Transducer Technology, Aerospace Information Research Institute, Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Xiaoye","middleName":"","lastName":"Huo","suffix":""},{"id":541403516,"identity":"35be889e-316e-4dba-94b7-04881ab67175","order_by":9,"name":"Ke Wang","email":"","orcid":"","institution":"School of Electronic Engineering, Beijing University of Posts and Telecommunications","correspondingAuthor":false,"prefix":"","firstName":"Ke","middleName":"","lastName":"Wang","suffix":""},{"id":541403517,"identity":"14b922fe-17ad-4ae0-ad65-5a88024680eb","order_by":10,"name":"Xianxiang Chen","email":"","orcid":"","institution":"Aerospace lnformation Research Institute,Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Xianxiang","middleName":"","lastName":"Chen","suffix":""},{"id":541403518,"identity":"46d88df4-446a-470f-a415-758b7d6c80bf","order_by":11,"name":"Jian Chen","email":"","orcid":"https://orcid.org/0000-0003-4612-3279","institution":"State Key Laboratory of Transducer Technology, Aerospace Information Research Institute, Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Jian","middleName":"","lastName":"Chen","suffix":""}],"badges":[],"createdAt":"2025-10-27 11:52:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7956597/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7956597/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96363697,"identity":"2a16b252-7ec1-4c7b-9a75-6c07a44d6194","added_by":"auto","created_at":"2025-11-20 10:07:46","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1934047,"visible":true,"origin":"","legend":"","description":"","filename":"20251025manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/1a9bac2b802b83d1d7908ee6.docx"},{"id":96283622,"identity":"546ea152-95d6-4eda-a021-0230145876f5","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13265,"visible":true,"origin":"","legend":"","description":"","filename":"MICRONANO05072.json","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/169e92b3dda88a6d311bb8f1.json"},{"id":96283623,"identity":"ffd573bf-edc3-4d23-8aac-74777bbffec4","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":155543,"visible":true,"origin":"","legend":"","description":"","filename":"InformedConsent.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/63cd990e1c6bf2df02ffb96a.pdf"},{"id":96364352,"identity":"d3ea96c7-2a9d-4497-8bbf-dd06599f62a1","added_by":"auto","created_at":"2025-11-20 10:09:13","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":101771,"visible":true,"origin":"","legend":"","description":"","filename":"MICRONANO050720enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/39dbb6f831513a6cfa0f129f.xml"},{"id":96283621,"identity":"f356335e-8e8d-4ba1-93ec-e39e6e9fcb3a","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":202976,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/6eda6854687e9af85cbab41a.jpeg"},{"id":96364183,"identity":"951d3ee2-3dd5-4c19-97ca-d3270abed9f0","added_by":"auto","created_at":"2025-11-20 10:09:01","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":993461,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/d11c549c6db059e9d8339729.jpeg"},{"id":96363645,"identity":"650d5d72-886d-441a-95ca-6cb5cac353a3","added_by":"auto","created_at":"2025-11-20 10:07:38","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":147590,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/e662a35108f40aa4b942e87f.jpeg"},{"id":96283627,"identity":"0992166a-06fa-4109-a6a4-eb2d8cfd7f34","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":484444,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/19e4e2feba0bb796727c120e.jpeg"},{"id":96283626,"identity":"b97483e3-b5d4-446e-bc4e-cbbd877ce121","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":38805,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/3ac48e35c95a155f0a25af12.png"},{"id":96363373,"identity":"0afd7e9a-2c07-4722-84d4-b0abea3d1a42","added_by":"auto","created_at":"2025-11-20 10:06:27","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":504450,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/f4051522e29520359978ebac.png"},{"id":96283629,"identity":"c2adaf79-8881-4953-be58-c31f874679c9","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23786,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/54ba7908e51637f315a15e2f.png"},{"id":96283628,"identity":"074d0059-2306-4b0a-9f90-40a6cf712bde","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":84215,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/90eb29bcdaaac42715391935.png"},{"id":96283630,"identity":"84c0067d-27f5-40d6-81b7-d20107bf106d","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"xml","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":100766,"visible":true,"origin":"","legend":"","description":"","filename":"MICRONANO050720structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/ee2ac401b09f6faf04c9f873.xml"},{"id":96283632,"identity":"a35a6255-5225-453e-90d7-0287ee7167b6","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"html","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":110542,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/f0d16d562449de07ea0b8188.html"},{"id":96283619,"identity":"57a59cf5-d053-4aac-9355-5157cdaf034c","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":516851,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic of generalizable self-supervised learning for imaging flow cytometry on multi-dataset leukocyte differential. (a) In image acquisition, images of leukocyte mixtures, purified leukocytes, and leukemia cell lines were captured by a home-developed imaging flow cytometry. Through multiple independent experiments, one non-annotated leukocyte mixture dataset along with three annotated datasets, which were utilized for testing model generalization, were constructed. (b) In self-supervised pre-training, a MoCo v2-based contrastive learning framework was conducted on the non-annotated leukocyte mixture dataset. This framework utilized a ResNet-18 as the shared backbone network, comprising a query encoder updated by backpropagation and a momentum encoder with momentum-updated parameters. After pre-training, the query encoder was saved as a feature extractor. (c) In downstream differential, the three annotated datasets were used independently to train and evaluate to validate the generalization. During fine-tuning, the pre-trained ResNet-18 weights were frozen, and only a lightweight linear classifier or MLP head was trained for each specific task.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/2f03bf08d980fdb5ee10a499.png"},{"id":96283615,"identity":"5228f034-7aa9-43b3-9d93-6dda7a92e0f3","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1693210,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative single-cell bright-field images collected by the home-developed imaging flow cytometry for the four datasets: (a) Non-annotated mixture of leukocytes, all of which are normal leukocytes; (b) Annotated 4-class normal leukocytes including eosinophils (EOS), lymphocytes (LYM), monocytes (MON), and neutrophils (NEU); (c) Annotated 2-class abnormal leukemia cell lines of HL60 and K562; (d) Annotated 8-class normal/abnormal leukocytes, comprising five types of normal leukocytes—basophils (BAS), eosinophils (EOS), lymphocytes (LYM), monocytes (MON), and neutrophils (NEU)—and three types of abnormal leukemia cell lines (Jurkat, HL60, and K562). (Scale bar: 5 μm)\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/652199a81347fc3f5933b04e.png"},{"id":96364629,"identity":"33a49dfb-08f5-4c4e-9f8a-c4e5a6315253","added_by":"auto","created_at":"2025-11-20 10:09:29","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":219531,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance comparison of MoCo v2 self-supervised pre-training under different learning rates. (a) Top-1 classification accuracies versus training epochs. (b) Contrastive losses versus training epochs. The learning rates were set to 0.01 (red), 0.02 (blue), 0.0225 (green), 0.025 (purple) and 0.05 (yellow). Results demonstrate that the performance of the three learning rates is remarkably similar, all achieving very high final accuracies of ~ 100% with losses all converging near 6.2, highlighting the parameter robustness of MoCo v2 on this dataset. The learning rate of 0.0225 (green curve) was ultimately selected as it achieved the highest accuracy of 99.99%.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/b4d0bda5c4d39aafd7186814.png"},{"id":96283618,"identity":"dcd8bd52-0a29-4308-9d9b-a0d2acecdc1c","added_by":"auto","created_at":"2025-11-19 11:49:34","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":796360,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrices for downstream leukocyte differential on three datasets: (a) the linear classification head and (b) the MLP head for 4-class normal leukocytes; (c) the linear classification head and (d) the MLP head for 2-class abnormalleukocytes; (e) the linear classification head and (f) the MLP head for 8-class normal/abnormal leukocytes. Evaluation shows that self-supervised pre-trained feature extractors were highly discriminative and generalizable, achieving high accuracy with linear evaluation alone; MLP heads further enhanced nonlinear fitting, boosting overall classification accuracy.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/ea948edf3e38406970c19909.png"},{"id":96369127,"identity":"9be93746-b2ec-40ae-b1e4-97c8449675bd","added_by":"auto","created_at":"2025-11-20 10:19:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4061896,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7956597/v1/307ab4ba-c454-4316-8e3b-baa360eb8f26.pdf"}],"financialInterests":"There is no conflict of interest","formattedTitle":"Generalizable Self-supervised Learning for Imaging Flow Cytometry on Multi-dataset Leukocyte Differential","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eLeukocyte differential refers to the quantification of the percentage and absolute counts of white blood cells (WBCs) in peripheral blood, and it is widely used as a first-line hematological test [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The gold-standard method for leukocyte differential analysis involves the microscopic examination of blood smears, in which a blood sample is applied to slides, stained with acidic eosin and basic methylene blue dyes, and observed under a microscope [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Based on the morphological analysis of nuclear characteristics, organelle architecture, and cytoplasmic or membrane properties, trained professionals in hematology are able to classify WBCs into recognized normal types\u0026mdash;including basophils (BAS), eosinophils (EOS), lymphocytes (LYM), monocytes (MON), and neutrophils (NEU)\u0026mdash;as well as various abnormal forms, such as myeloblasts, promyelocytes, myelocytes, metamyelocytes, and atypical lymphocytes [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Although widely used in clinical practice, this method is time-consuming, which seriously limits diagnostic efficiency [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTo address this issue, hematology analyzers were designed to perform continuous analysis of single blood cells based on light scattering and impedance variations, enabling high-throughput leukocytes differentiation [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Automated commercial hematology analyzers, such as the Sysmex XN-2000 and Beckman Coulter DxH 900, are routinely utilized in clinical laboratories worldwide [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Unfortunately, accurate leukocyte differential was achieved only for normal WBCs, but not for abnormal ones, due to the absence of morphological parameters [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTo overcome these limitations, imaging flow cytometry was utilized to capture fluorescence and/or bright-field images of white blood cells for classification by combining machine learning and deep learning [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. For example, the commercial ImageStream platform attained 97.0% F1-scores in four-part leukocyte classification with machine learning [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. In 2024, Wang, et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] accomplished a classification accuracy of 97.1% for three ovarian cell types by integrating three supervised networks on brightfield and three-channel fluorescence imaging data. In 2025, Chen, et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] realized a classification accuracy of 99.6% for four leukocyte subtypes through five deep neural network fusion strategies of supervised LSTM and ResNet-18 applied to impedance and fluorescence imaging data. However, since supervised learning relies on annotated cell images and this labeling process generates cell losses, current imaging flow cytometry with supervised learning cannot be used for leukocyte differential.\u003c/p\u003e\u003cp\u003eThus, label-free bright-field images captured by imaging flow cytometry [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] coupled with self-supervised learning [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] holds a significant promise for leukocyte differential. Different from supervised learning, self-supervised learning can utilize unlabeled data for training and classification by automatically generating proxy tasks to generate supervisory signals from the data itself, thereby obtaining guidance from large-scale data and achieving highly discriminative features for downstream classification [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. Within the field of leukocyte classification of blood smears, self-supervised learning has demonstrated high accuracies and strong generalization. For example, in 2024, Jia, et al. [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] achieved a classification accuracy of 93.51% across 15 classes on the TMAMD leukocyte-stained blood smear dataset using a masked autoencoder pretrained on a stained blood smear dataset. In 2024, Koch, et al. [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] reported a classification accuracy of 86.6% for 10-part leukocyte differential on the Acevedo dataset, utilizing a pretrained DINO v2 network.\u003c/p\u003e\u003cp\u003eAs to leukocyte classification based on imaging flow cytometry and self-supervised learning, the only previous report was in 2023, in which Zhang, et al. [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] reported an accuracy of 91.9% for three-part classification of normal leukocytes using a pretrained deep convolutional autoencoder with frozen parameters. However, the generative learning used in this study focused on pixel-level reconstruction, which over-emphasized low-level details irrelevant to semantic discrimination, thereby potentially compromising the accuracy and generalizability of leukocyte classification.\u003c/p\u003e\u003cp\u003eThis study establishes a self-supervised contrastive framework for leukocyte differential, based on bright-field images from imaging flow cytometry. Compared to generative learning, contrastive learning directly enhances feature separability and robustness against perturbations without being constrained by the need for precise input reconstruction. More specifically, the feature extraction network was pretrained on unannotated leukocyte images and then used for three downstream leukocyte differential tasks, including those involving previously unseen subtypes with lightweight linear or MLP classifiers, while freezing the feature extractor parameters. The MoCo v2-based methodology achieves performance parity with supervised baselines, underscoring exceptional cross-dataset generalization capabilities without architectural retraining. The unannotated and strongly generalizable imaging flow cytometry-based sorting method may enhance the capability of clinical leukocyte differential.\u003c/p\u003e"},{"header":"2. Materials and methodology","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Working flow chart\u003c/h2\u003e\u003cp\u003eThe overall workflow of this study, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, comprised three major phases: (a) image acquisition, (b) self-supervised pre-training, and (c) muti-dataset downstream leukocyte differential. Initially, leukocyte samples\u0026mdash;including mixed leukocytes, purified leukocytes, and leukemia cell lines\u0026mdash;were captured using a home-developed imaging flow cytometry system. Through multiple independent experiments, one non-annotated dataset (derived from leukocyte mixtures) and three annotated datasets were constructed for pre-training and generalization evaluation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn the self-supervised pre-training phase, a Momentum Contrast (MoCo v2) framework [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e] was employed to learn representative features from the non-annotated dataset. This framework incorporated a ResNet-18 [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e] architecture as the shared backbone network, consisting of a query encoder (updated via backpropagation) and a momentum encoder (updated via momentum-based moving average). The pre-trained query encoder was subsequently saved as a fixed feature extractor.\u003c/p\u003e\u003cp\u003eFor the downstream differential phase, the pre-trained weights were frozen, and only a lightweight linear classifier or multilayer perceptron (MLP) head was trained and evaluated on each of the three annotated datasets to validate cross-dataset generalization.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e\u003cem\u003e2.2 Image acquisition and dataset composition\u003c/em\u003e\u003c/h2\u003e\u003cp\u003eThe image data in this study were acquired by a home-developed imaging flow cytometry (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e (a)). The cytometry utilized a PZT-based acoustic focusing method, which ensured that cells pass sequentially through the depth-of-field region at the center of the channel, enabling high-throughput acquisition of high-quality, in-focus bright-field cell images. The samples tested included leukocyte mixtures, purified leukocytes, and leukemia cell lines. For detailed configurations regarding materials, cell preparation, and the imaging flow cytometry setup, please refer to our group\u0026rsquo;s previous publication [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThrough multiple experiments, four mutually independent datasets were constructed: one non-annotated dataset for self-supervised pre-training and three annotated datasets for downstream differential tasks to test generalization of pre-training feature exactor. No cell images were reused across these datasets. The self-supervised pre-training dataset consisted of images from leukocyte mixtures obtained from four healthy volunteers and contained only non-annotated images of normal leukocytes. The three downstream differential datasets all comprised annotated leukocyte images. Normal leukocyte types came from purified leukocytes isolated via fluorescence-activated cell sorting after red blood cell lysis of the whole blood, while abnormal leukocyte types were derived from in vitro cultured leukemia cell lines. All three datasets are independent and non-overlapping.\u003c/p\u003e\u003cp\u003eThe first dataset included the four most abundant normal leukocyte types\u0026mdash;EOS (eosinophil), LYM (lymphocyte), MON (monocyte), and NEU (neutrophil)\u0026mdash;and served as a four-class baseline for normal cell classification. The second dataset contained two types of leukemia cell lines, HL60 (promyelocytic leukemia) and K562 (chronic myelogenous leukemia), which were unseen during feature extractor pre-training. It was used to validate the generalizability of the feature exactor in classifying novel abnormal cell types. The third dataset comprised five types of normal leukocytes\u0026mdash;BAS (basophil), EOS, LYM, MON, and NEU\u0026mdash;as well as three types of abnormal leukemia cell lines: Jurkat (T-cell leukemia), HL60, and K562. This dataset included both all categories of normal cells and abnormal cell types (representing both lymphoid and myeloid lineages) not encountered during feature extractor pre-training.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Self-supervised Pre-training\u003c/h2\u003e\u003cp\u003eThe self-supervised learning workflow in this study consisted of two parts. First, a feature extractor was pre-trained using self-supervised learning on a large amount of non-annotated data. Then, for the three downstream differential tasks, the deep neural network for feature extraction was frozen, and only a simple classifier based on the extracted features was trained separately using the annotated datasets. This approach completed cell classification with minimal training cost while keeping the backbone feature extraction deep neural network unchanged. The training in this study was performed on Ubuntu 18.04 systems using PyTorch 1.8.2 and CUDA 11.1, hosted on AutoDL servers with an Intel Xeon Gold 6430 CPU (16 vCPUs) and a single NVIDIA RTX 4090 GPU (24 GB).\u003c/p\u003e\u003cp\u003eIn the self-supervised pre-training phase, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e (b), this study employed the MoCo v2 architecture for self-supervised contrastive learning. Its key features included a momentum encoder and a queue dictionary for maintaining a large set of negative samples. The backbone network for feature extraction was the convolutional neural network ResNet-18, which offers a good balance between model complexity and accuracy, effectively meeting the feature extraction requirements for leukocyte images. Since the effectiveness of contrastive learning depended on the data augmentation strategy, this study designed a composite augmentation scheme including rotation, flipping, erosion, and Gaussian blur, tailored to the characteristics of non-annotated bright-field leukocyte images, which were single-channel and sensitive to magnification.\u003c/p\u003e\u003cp\u003eThe batch size was set to 256, the maximum feasible on the RTX 4090 GPU (24 GB), to ensure high training speed. Key hyperparameters, including learning rates of 0.01 and 0.05, then intermediate values of 0.02 and 0.025, and finally 0.0225, were tuned to identify the optimal network for downstream feature extraction. The training performance of MoCo v2 was evaluated using two primary metrics: Top-1 accuracy, where higher values indicated better representation quality upon convergence, and contrastive loss, where lower values signified more effective learning of invariant features during pre-training until convergence was achieved.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Multi-dataset Downstream Leukocyte Differential\u003c/h2\u003e\u003cp\u003eIn the three downstream leukocyte differential tasks, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e (c), based on the feature extractor pre-trained with the aforementioned MoCo v2 self-supervised approach, two schemes were employed for training (with 5-fold cross-validation): a single linear classification head and a 3-layer MLP. In the linear head scheme, the feature extractor was frozen, and only a single layer (the linear head) was trained. Since no new features were introduced into the feature space, only linear division of existing features occurred, its classification results could demonstrate the effectiveness of the features extracted by self-supervised learning. Introducing a 3-layer MLP, which froze the feature extraction network and used only a simple non-linear classification network for fine-tuning, could improve the classification accuracy of downstream tasks based on features extracted by self-supervised learning with minimal training cost compared to retraining the whole deep neural networks.\u003c/p\u003e\u003cp\u003eTo contrast the effectiveness of the above scheme, this study also employed another self-supervised learning scheme and three supervised learning schemes for comparison on the three downstream tasks. All tasks used 5-fold cross-validation, and key hyperparameters like learning rate were tuned for optimal results. A batch size of 256 was used for the fastest training speed on the available GPU.\u003c/p\u003e\u003cp\u003eSimCLR was used as another self-supervised scheme. Its effectiveness was measured by freezing the feature-extracting ResNet-18 and training the same linear classification head on the three downstream tasks, comparing it to the feature extraction network trained with MoCo v2 to demonstrate differences in feature effectiveness among self-supervised schemes. The SimCLR self-supervised feature extractor training also used ResNet-18 as the backbone network with the same data augmentation as MoCo v2. However, as SimCLR lacks momentum update and its effectiveness relies on a large batch size, and the computational power of the single RTX 4090 (24 GB) was limited, gradient accumulation was used. The original batch size was 256 (the maximum on the RTX 4090), with a gradient accumulation of 16, resulting in an effective batch size of 4096 to meet SimCLR's training requirements. Other key hyperparameters included a temperature of 0.5 and a learning rate of 0.1. For the downstream linear classification head training, the ResNet-18 backbone network trained by SimCLR was frozen, and only the single neural network layer was trained, exactly the same as the linear classification head used with the MoCo v2-trained feature extractor.\u003c/p\u003e\u003cp\u003eThe supervised learning methods compared on the three downstream tasks were trained separately using the annotated datasets from the respective downstream task. The employed supervised learning methods included MLP, ResNet-18, and ViT-small.\u003c/p\u003e\u003cp\u003eMLP is a multilayer perceptron, a feedforward artificial neural network model composed of an input layer, hidden layers, and an output layer, capable of learning complex patterns and relationships in data. Using only MLP for supervised learning allowed comparison with using the self-supervised feature extractor followed by MLP, thus demonstrating the effectiveness of the self-supervised learned features.\u003c/p\u003e\u003cp\u003eResNet-18, as a representative convolutional neural network with balanced performance, is an 18-layer deep residual network that effectively alleviates the vanishing gradient problem in deep networks through skip connections. It was a lightweight convolutional neural network that achieves high accuracy with a relatively small number of model parameters and served as a comparison for the supervised learning convolutional neural network architecture against the scheme in this study.\u003c/p\u003e\u003cp\u003eVision Transformer (ViT), proposed by Google, is a Transformer-based network for computer vision. It divides images into fixed-size patches and treats them as token sequences, extracting features through a Transformer encoder to handle visual tasks. ViT-small, as the smallest model, had a number of network parameters roughly on par with ResNet-18 and represented the supervised learning ViT architecture compared to the scheme in this study.\u003c/p\u003e\u003cp\u003eIn the comparison of results from the three downstream tasks, classification accuracy, precision, recall, and F1-score were calculated for evaluation. Accuracy measured the proportion of correctly predicted samples. Precision focused on the proportion of true positive samples among those predicted as positive. Recall focused on the proportion of correctly predicted samples among all true positive samples. F1-score was the harmonic mean of precision and recall, providing a comprehensive assessment of model performance.\u003c/p\u003e\u003cp\u003eFor the core scheme in this study, the self-supervised scheme based on MoCo v2, its results on the three downstream tasks using both the linear classification head and the 3-layer MLP classification head were additionally evaluated using confusion matrices. The confusion matrix is a tool that visualizes the correspondence between predicted results and true samples in a tabular format, used for detailed analysis of the types of classification errors. Each cell within the matrix contains two values: the absolute count of instances and a corresponding proportion. The proportion is calculated by taking the number of cases in a given cell and dividing it by the total number of instances associated with the true class of that row, thereby providing a normalized view of the prediction distribution for each actual class.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results and discussion","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Datasets\u003c/h2\u003e\u003cp\u003eThe details of the image datasets captured by the home-developed imaging flow cytometry were shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. It started with the leukocyte mixture images used for the self-supervised feature extraction task. These were sourced from 4 healthy donors, with a total of 80,347 images. This portion of cell images was obtained directly from the leukocyte mixture after the lysis of red blood cells in whole blood. They were non-annotated, used for training a stable and effective feature extractor in the self-supervised task. Representative images were shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (a).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eLeukocyte datasets captured by a home-developed imaging flow cytometry.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDataset Name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePrimary Use\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAnnotate Status\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTotal\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eCell Classes \u0026amp; Quantity\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLeukocyte Mixture Dataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSelf-supervised Pre-training\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNon-annotated\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e80,347\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eNormal Leukocytes: 80,347\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003e4-class Normal Leukocyte Dataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eTask 1: 4-class Normal Leukocyte Differential\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eAnnotated\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003e24,834\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1. Eosinophil (EOS): 4,175\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2. Lymphocyte (LYM): 5,879\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3. Monocyte (MON): 6,383\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4. Neutrophil (NEU): 8,397\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003e2-class Abnormal Leukocyte Dataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eTask 2: 2-class Abnormal Leukocyte Differential\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eAnnotated\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003e11,093\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1. HL60: 5,468\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2. K562: 5,625\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e\u003cp\u003e8-class Normal/Abnormal Leukocyte Dataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\" morerows=\"7\" rowspan=\"8\"\u003e\u003cp\u003eTask 3: 8-class Normal/Abnormal Leukocyte Differential\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\" morerows=\"7\" rowspan=\"8\"\u003e\u003cp\u003eAnnotated\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\" morerows=\"7\" rowspan=\"8\"\u003e\u003cp\u003e8,419\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1. Basophil (BAS): 515\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2. Eosinophil (EOS): 1,034\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3. Lymphocyte (LYM): 1,159\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4. Monocyte (MON): 1,189\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5. Neutrophil (NEU): 1,238\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e6. Jurkat: 1,086\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e7. HL60: 1,061\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e8. K562: 1,137\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThen, annotated leukocyte image data were captured and subsequently used in the three downstream differential tasks. The normal leukocyte types were derived from purified leukocytes, which were obtained via fluorescence-activated cell sorting (FACS) after red blood cell lysis in whole blood from healthy donors. The abnormal leukocyte types were derived from in vitro cultured leukemia cell lines. The three datasets were based on mutually independent data with no overlap. Representative images were shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (b) \u0026ndash; (d). Dataset 1 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (b)) contained 4 types of purified normal leukocytes: EOS (eosinophils), LYM (lymphocytes), MON (monocytes), and NEU (neutrophils), all from human peripheral blood of healthy individuals. The total data volume was relatively large at 24,834. Dataset 2 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (c)) contained 2 types of leukemia cells: HL60 and K562, both of which were abnormal cell types not present during the training of the self-supervised learning feature extractor. The total data volume was moderate at 11,093. Dataset 3 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (d)) contained 8 types of normal and abnormal leukocytes: 5 types of purified normal leukocytes (BAS (basophils), EOS (eosinophils), LYM (lymphocytes), MON (monocytes), NEU (neutrophils)) and 3 types of leukemia cells (HL60, Jurkat, and K562). These three leukemia cell types were also not present during the training of the self-supervised learning feature extractor. The total data volume was the smallest at 8,419, but it had the most categories. Except for BAS (basophils), which were fewer at 515, the other categories had around 1,000 images each to ensure relatively balanced data volume.\u003c/p\u003e\u003cp\u003eJudging from the example images, they were all clear leukocyte bright-field images without defocusing, and the cell contours and internal textures were distinct. Based on the visually apparent cell size, normal leukocytes were smaller than abnormal leukemia cells, while the size of normal leukocytes remained consistent across datasets. Among the normal leukocytes, lymphocytes were significantly smaller than the other four types, which was consistent with standard leukocyte analysis. The leukocyte mixture contained cells with both larger and smaller diameters, suggesting that it included all categories of normal leukocytes.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Results of Self-supervised Pre-training\u003c/h2\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presented the self-supervised pre-training results on the non-annotated leukocyte mixture dataset, specifically comparing the performance of MoCo v2 under different learning rates. This included the variation of Top-1 accuracy with training epochs (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e (a)) and the variation of contrastive loss with training epochs (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e (b)). The learning rates tested were 0.01 (red), 0.02 (blue), 0.0225 (green), 0.025 (purple), and 0.05 (yellow). These results indicated that the model converged around 100 epochs under all five learning rates, justifying the selection of 200 epochs as the training duration.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAmong the five learning rates, the network selected for feature extraction in downstream tasks was chosen based on the highest Top-1 accuracy. The final selection was the model trained with a learning rate of 0.0225 (green curve), as it achieved the best Top-1 accuracy (99.99%), outperforming the 99.97% achieved with 0.01, 99.98% with 0.02, 99.98% with 0.025, and the 99.94% achieved with 0.05. Regarding the contrastive loss, since this study used a self-constructed dataset, the final converged contrastive loss value of approximately 6.2 only indicated effective convergence of self-supervised learning on this specific dataset and was not directly comparable to contrastive loss values reported on other datasets.\u003c/p\u003e\u003cp\u003eNote that the final performance in terms of both Top-1 accuracy and contrastive loss were highly similar across the learning rates, achieving an extremely high final accuracy (close to 100%) and a contrastive loss value approaching 6.2. This demonstrated the parameter robustness of MoCo v2 on this dataset and confirmed the high compatibility of the selected data augmentation strategy, the ResNet-18 backbone network, and the leukocyte mixture dataset, marking a successful phase in self-supervised pre-training.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Results of Multi-dataset Downstream Leukocyte Differential\u003c/h2\u003e\u003cp\u003eAfter obtaining a feature extraction network through self-supervised pre-training, the network was frozen and applied to differential tasks on three downstream leukocyte datasets. For each dataset, a linear classification head was first trained to validate the effectiveness of the features, followed by training an MLP classification head to improve classification accuracy. The classification confusion matrix results were shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. In the 4-class normal leukocyte differential task (see Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e (a)), the linear classification head achieved high precision with each cell type recall exceeding 0.9, demonstrating the effectiveness of feature extraction. The MLP classification head further improved the recall for each cell type to above 0.94 (see Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e (b)). In terms of cell types, the two lowest recall values were for EOS and NEU, which aligns with the expectation that eosinophils and neutrophils are similar and thus difficult to distinguish.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn the 2-class abnormal leukemia cell differential task (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(c) - (d)), the model exhibited exceptional accuracy in distinguishing between HL60 and K562, two types of abnormal leukocytes not present in the pre-training dataset, with each class recall exceeding 0.95, effectively proving the generalization capability of the self-supervised pre-trained feature extraction network. The MLP classification head effectively reduced the minor misclassifications observed with the linear classification head, achieving a class-wise recall exceeding 0.97. The high accuracy might be attributed to the relative simplicity of the binary classification task.\u003c/p\u003e\u003cp\u003eIn the most complex 8-class normal/abnormal mixed leukocyte differential task (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e (e) \u0026ndash; (f)), the linear classification head still achieved a minimum recall of 0.84 for NEU, more robustly demonstrating the generalization of the feature extraction network. The recall for each cell type was relatively lower due to the task's complexity. In the confusion matrix for linear classification shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e (e), LYM exhibited the highest recall of 0.96, likely because their significantly smaller diameter makes them a distinctive feature. They were followed by three types of abnormal leukocytes, which had larger diameters than normal leukocytes. The remaining four types of normal leukocytes, with similar diameters, showed lower recall. These results were consistent with cell diameter being the most obvious and physically meaningful distinguishing characteristic. The total number of misclassified samples between normal and abnormal cells was 146, which was less than the total number of misclassified samples within normal leukocytes (573), verifying that greater feature differences exist between normal and abnormal cells.\u003c/p\u003e\u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e (f), the MLP classification head significantly improved the recall for several leukocyte types (e.g., 0.90 of MON, 0.87 of NEU, and 0.96 of Jurkat). However, the improvement for BAS was limited, with only two additional correctly classified samples, likely because the number of BAS samples was only about half that of other leukocytes and the MLP classification head had a simple structure with limited capacity. Despite this, high-accuracy differential with each leukocyte type recall exceeding 85% was achieved.\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presented the performance of the proposed self-supervised learning scheme compared to other different supervised and self-supervised methods on three leukocyte datasets, evaluating Accuracy, Precision, Recall, and F1-score to comprehensively showcase model performance. First, the supervised learning baseline, which involved direct MLP classification on the data, performed significantly lower than other methods, with accuracies of only 43.40% and 43.37% for the 4-class Normal and 8-class Normal/Abnormal tasks, respectively. This demonstrated that simply using an MLP or even simpler linear layers could not effectively distinguish leukocytes, thus highlighting the necessity and effectiveness of feature extraction through self-supervised pre-training when compared to training classification heads afterwards.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eClassification performance of different supervised/self-supervised methods on three leukocyte datasets.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDataset\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMethodology\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eF1-score\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e\u003cp\u003e4-class Normal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMLP (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e43.40%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e45.33%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e36.04%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e33.00%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eResNet-18 (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e96.81%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e96.80%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e96.79%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e96.78%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eViT-Small (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e97.61%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e97.63%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.47%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.54%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSimCLR (Linear Head)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e73.58%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e72.96%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e72.68%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e72.81%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eMoCo v2 (Linear Head)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e93.56%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e93.70%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e93.45%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e93.57%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eMoCo v2 (MLP Head)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e96.16%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e96.15%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e95.98%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e96.06%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e\u003cp\u003e2-class Abnormal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMLP (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e85.28%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e85.29%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e85.26%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e85.27%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eResNet-18 (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e97.67%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e97.68%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.67%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.67%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eViT-Small (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e95.17%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e95.17%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e95.16%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e95.17%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSimCLR (Linear Head)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e95.46%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e95.45%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e95.46%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e95.46%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eMoCo v2 (Linear Head)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e95.47%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e95.46%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e95.47%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e95.46%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eMoCo v2 (MLP Head)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e96.68%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e96.69%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e96.68%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e96.68%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e\u003cp\u003e8-class Normal/\u003c/p\u003e\u003cp\u003eAbnormal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMLP (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e43.37%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e43.13%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e40.34%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e39.63%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eResNet-18 (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e94.01%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e93.83%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e94.11%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e93.88%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eViT-Small (Sup)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e92.54%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e92.49%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e92.25%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e92.37%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSimCLR (Linear Head)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e72.70%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e71.87%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e71.40%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e71.46%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eMoCo v2 (Linear Head)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e91.46%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e91.12%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e91.15%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e91.13%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eMoCo v2 (MLP Head)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e92.24%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e92.00%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e91.87%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e91.92%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eSubsequently, two representative supervised learning networks, ResNet-18 and ViT-Small, provided performance benchmarks on the three datasets. It was important to note that supervised learning required retraining the entire feature extraction network separately on each annotated dataset and could not leverage non-annotated data. Then, two self-supervised learning contrastive methods, SimCLR and MoCo v2, were evaluated based on linear classification head performance. MoCo v2 outperformed SimCLR by approximately 20% across all four metrics (Accuracy, Precision, Recall, F1-score) on both the 4-class Normal and 8-class Normal/Abnormal tasks, proving that the features learned by MoCo v2 through self-supervised learning possess higher discriminability and generalization. Finally, by freezing the feature extraction network pre-trained with MoCo v2 and training only a simple MLP classification head, the achieved performance on all four metrics was within approximately 2% of the supervised learning benchmarks across all three downstream differential tasks. This fully demonstrated that the self-supervised pre-trained feature extraction network possesses exceptional cross-dataset generalization capability.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eThis study establishes a self-supervised contrastive framework for leukocyte differential, based on bright-field images from imaging flow cytometry. By pretraining on non-annotated data, the feature extractor learned discriminative representations invariant to augmentation, achieving high performance on downstream tasks with minimal fine-tuning. The method generalizes across diverse datasets, including unseen abnormal leukocyte types, with accuracy rivaling supervised models. This underscores the potential of self-supervised learning to enhance flow cytometry by reducing manual annotation burdens and improving model generalization capability.\u003c/p\u003e\u003cp\u003eFuture work will extend to broader representative patient samples, classification of rare cell populations, and multimodal diagnostic integration including fluorescence labeling or impedance signals, promoting the clinical application of intelligent diagnosis in flow cytometry.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eCompeting interests\u003c/h2\u003e\u003cp\u003eThe authors declare no conflicts of interest.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThis work was supported from the National Natural Science Foundation of China (Grants No.: 62331025 and 62121003) and the Beijing Municipal Natural Science Foundation (Grants No.: 4232010).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eJ. A. Molnar, \"12 - Automated blood cell analysis,\" \u003cem\u003eRodak's Hematology (Sixth Edition)\u003c/em\u003e, E. M. Keohane, C. N. Otto and J. M. Walenga, eds., pp. 174\u0026ndash;200, St. Louis (MO): Elsevier, 2020.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJ. E. Lewis, and O. Pozdnyakova, \u0026ldquo;Digital assessment of peripheral blood and bone marrow aspirate smears,\u0026rdquo; \u003cem\u003eInternational Journal of Laboratory Hematology\u003c/em\u003e, vol. 45, pp. 50\u0026ndash;58, Jun. 2023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eM. L. Chase et al., \u0026ldquo;Consensus recommendations on peripheral blood smear review: Defining curricular standards and fellow competency,\u0026rdquo; \u003cem\u003eBlood Advances\u003c/em\u003e, vol. 7, no. 13, pp. 3244\u0026ndash;3252, Jul. 2023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eM. O. Nakashima, S. N. Coulter, B. J. Blond, R. W. Brown, and J. A. Vos, \u0026ldquo;The impact of pathologist review on peripheral blood smears a college of american pathologists q-probes study of 22 laboratories,\u0026rdquo; \u003cem\u003eArchives of Pathology \u0026amp; Laboratory Medicine\u003c/em\u003e, vol. 149, no. 7, pp. 618\u0026ndash;626, Jul. 2025.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eL. Palmer et al., \u0026ldquo;ICSH recommendations for the standardization of nomenclature and grading of peripheral blood cell morphological features,\u0026rdquo; \u003cem\u003eInternational Journal of Laboratory Hematology\u003c/em\u003e, vol. 37, no. 3, pp. 287\u0026ndash;303, Jun. 2015.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eC. Matek, S. Schwarz, K. Spiekermann, and C. Marr, \u0026ldquo;Human-level recognition of blast cells in acute myeloid leukaemia with convolutional neural networks,\u0026rdquo; \u003cem\u003eNature Machine Intelligence\u003c/em\u003e, vol. 1, no. 11, pp. 538\u0026ndash;544, Nov. 2019.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eC. Vilchez, \"13 - Examination of the peripheral blood film and correlation with the complete blood count,\" \u003cem\u003eRodak's Hematology (Sixth Edition)\u003c/em\u003e, E. M. Keohane, C. N. Otto and J. M. Walenga, eds., pp. 201\u0026ndash;218, St. Louis (MO): Elsevier, 2020.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZ. X. Tian, Y. C. Wei, Y. L. Yu, F. L. Zhou, and Z. L. Huang, \u0026ldquo;Blood cell analysis: From traditional methods to super-resolution microscopy,\u0026rdquo; \u003cem\u003ePhotonics\u003c/em\u003e, vol. 9, no. 4, pp. 261, Apr. 2022.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eW. X. J. Cheng et al., \u0026ldquo;Application of image recognition technology in pathological diagnosis of blood smears,\u0026rdquo; \u003cem\u003eClinical and Experimental Medicine\u003c/em\u003e, vol. 24, no. 1, pp. 181, Aug. 2024.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eS. Rastogi et al., \u0026ldquo;Overall equipment effectiveness, efficiency and slide review analysis of high-end hematology analyzers,\u0026rdquo; \u003cem\u003ePractical Laboratory Medicine\u003c/em\u003e, vol. 30, pp. e00275, May. 2022.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eM. Daves et al., \u0026ldquo;Modern hematology analyzers: Beyond the simple blood cells count (with focus on the red blood cells),\u0026rdquo; \u003cem\u003eJournal of Laboratory and Precision Medicine\u003c/em\u003e, vol. 9, pp. 4, Jan. 2024.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eL. A. Liu et al., \u0026ldquo;Multi-wavelength multi-direction laser light scattering for cell characterization using machine learning-based methods,\u0026rdquo; \u003cem\u003eCytometry Part A\u003c/em\u003e, vol. 103, no. 10, pp. 796\u0026ndash;806, Oct. 2023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eS. Buoro et al., \u0026ldquo;Multicenter evaluation of blood cell counts on 8 automated hematology analyzers,\u0026rdquo; \u003cem\u003eBiochimica Clinica\u003c/em\u003e, vol. 40, no. 3, pp. 195\u0026ndash;203, Sep. 2016.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eO. Ciepiela et al., \u0026ldquo;A comparison of Mindray BC-6800, Sysmex XN-2000, and Beckman Coulter LH750 automated hematology analyzers: A pediatric study,\u0026rdquo; \u003cem\u003eJournal of Clinical Laboratory Analysis\u003c/em\u003e, vol. 30, no. 6, pp. 1128\u0026ndash;1134, Nov. 2016.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eL. Meintker, J. Ringwald, M. Rauh, and S. W. Krause, \u0026ldquo;Comparison of automated differential blood cell counts from Abbott Sapphire, Siemens Advia 120, Beckman Coulter DxH 800, and Sysmex XE-2100 in normal and pathologic samples,\u0026rdquo; \u003cem\u003eAmerican Journal of Clinical Pathology\u003c/em\u003e, vol. 139, no. 5, pp. 641\u0026ndash;650, May. 2013.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eE. Schapkaitz, and S. Raburabu, \u0026ldquo;Performance evaluation of the new measurement channels on the automated Sysmex XN-9000 hematology analyzer,\u0026rdquo; \u003cem\u003eClinical Biochemistry\u003c/em\u003e, vol. 53, pp. 132\u0026ndash;138, Mar. 2018.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eM. R. Wang et al., \u0026ldquo;Developments of conventional and microfluidic flow cytometry enabling high-throughput characterization of single cells,\u0026rdquo; \u003cem\u003eBiosensors-Basel\u003c/em\u003e, vol. 12, no. 7, pp. 443, Jul. 2022.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQ. Huang et al., \u0026ldquo;Imaging flow cytometry: From high - resolution morphological imaging to innovation in high - throughput multidimensional biomedical analysis,\u0026rdquo; \u003cem\u003eFrontiers in Bioengineering and Biotechnology\u003c/em\u003e, vol. 13, pp. 12, May. 2025.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eV. Railean, and B. Buszewski, \u0026ldquo;Flow cytometry - sophisticated tool for basic research or/and routine diagnosis; impact of the complementarity in both pre- as well as clinical studies,\u0026rdquo; \u003cem\u003eCritical Reviews in Analytical Chemistry\u003c/em\u003e, vol. 54, no. 7, pp. 2087\u0026ndash;2109, Oct. 2024.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eD. Vonficht et al., \u0026ldquo;Ultra-high-scale cytometry-based cellular interaction mapping,\u0026rdquo; \u003cem\u003eNature Methods\u003c/em\u003e, vol. 22, no. 9, Aug. 2025.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eH. Kanno et al., \u0026ldquo;High-throughput fluorescence lifetime imaging flow cytometry,\u0026rdquo; \u003cem\u003eNature Communications\u003c/em\u003e, vol. 15, no. 1, pp. 7376, Sep. 2024.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eM. Nassar et al., \u0026ldquo;Label-free identification of white blood cells using machine learning,\u0026rdquo; \u003cem\u003eCytometry Part A\u003c/em\u003e, vol. 95, no. 8, pp. 836\u0026ndash;842, May. 2019.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZ. W. Wang, Q. Liu, J. Zhou, and X. T. Su, \u0026ldquo;Single-detector multiplex imaging flow cytometry for cancer cell classification with deep learning,\u0026rdquo; \u003cem\u003eCytometry Part A\u003c/em\u003e, vol. 105, no. 9, pp. 666\u0026ndash;676, Sep. 2024.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eS. Chen et al., \u0026ldquo;Bimodal fusion of single-cell impedance and imaging data using five deep neural networks,\u0026rdquo; \u003cem\u003eIEEE Nanotechnology Magazine\u003c/em\u003e, vol. 19, no. 1, pp. 24\u0026ndash;30, Jan. 2025.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZ. Q. Du et al., \u0026ldquo;Label-free detection and enumeration of rare circulating tumor cells by bright-field image cytometry and multi-frame image correlation analysis,\u0026rdquo; \u003cem\u003eLab on a Chip\u003c/em\u003e, vol. 22, no. 18, pp. 3390\u0026ndash;3401, Sep. 2022.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eC. F. Otesteanu et al., \u0026ldquo;A weakly supervised deep learning approach for label-free imaging flow-cytometry-based blood diagnostics,\u0026rdquo; \u003cem\u003eCell Reports Methods\u003c/em\u003e, vol. 1, no. 6, pp. 100094, Oct. 2021.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZ. W. Wang, Q. Liu, R. Chu, K. Song, and X. T. Su, \u0026ldquo;Single-detector dual-modality imaging flow cytometry for label-free cell analysis with machine learning,\u0026rdquo; \u003cem\u003eOptics and Lasers in Engineering\u003c/em\u003e, vol. 168, pp. 107665, Sep. 2023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eL. L. Zeune et al., \u0026ldquo;Deep learning of circulating tumour cells,\u0026rdquo; \u003cem\u003eNature Machine Intelligence\u003c/em\u003e, vol. 2, no. 2, pp. 124\u0026ndash;133, Feb. 2020.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eB. Liu et al., \u0026ldquo;Self-supervised learning reveals clinically relevant histomorphological patterns for therapeutic strategies in colon cancer,\u0026rdquo; \u003cem\u003eNature Communications\u003c/em\u003e, vol. 16, no. 1, pp. 2328, Mar. 2025.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eT. Hong, M. Peng, Y. Kim, H. E. Schellhorn, and Q. Fang, \u0026ldquo;Automated cell profiling in imaging flow cytometry with annotation-efficient learning,\u0026rdquo; \u003cem\u003eOptics \u0026amp; Laser Technology\u003c/em\u003e, vol. 181, pp. 111992. 2025.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eY. N. Lu, L. J. Wen, J. Z. Liu, Y. J. Liu, and X. M. Tian, \u0026ldquo;Self-supervision can be a good few-shot learner,\u0026rdquo; in \u003cem\u003eProc. Computer Vision (ECCV 2022)\u003c/em\u003e, Tel Aviv, Israel, Nov. 2022, pp. 740\u0026ndash;758.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eV. Rani, S. T. Nabi, M. Kumar, A. Mittal, and K. Kumar, \u0026ldquo;Self-supervised Learning: A Succinct Review,\u0026rdquo; \u003cem\u003eArchives of Computational Methods in Engineering\u003c/em\u003e, vol. 30, no. 4, pp. 2761\u0026ndash;2775, May. 2023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eN. Jia et al., \u0026ldquo;A fine-grained image classification algorithm based on self-supervised learning and multi-feature fusion of blood cells,\u0026rdquo; \u003cem\u003eScientific Reports\u003c/em\u003e, vol. 14, no. 1, pp. 22964, Oct. 2024.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eV. Koch et al., \u0026ldquo;DinoBloom: A Foundation Model for Generalizable Cell Embeddings in Hematology,\u0026rdquo; in \u003cem\u003eProc. Medical Image Computing and Computer Assisted Intervention (MICCAI 2024)\u003c/em\u003e, Marrakesh, Morocco. 2024, pp. 520\u0026ndash;530.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZ. Zhang et al., \u0026ldquo;Interpretable unsupervised learning enables accurate clustering with high-throughput imaging flow cytometry,\u0026rdquo; \u003cem\u003eScientific Reports\u003c/em\u003e, vol. 13, no. 1, pp. 20533, Nov. 2023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eK. M. He et al., \u0026ldquo;Momentum contrast for unsupervised visual representation learning,\u0026rdquo; in \u003cem\u003eProc.\u003c/em\u003e 2020 \u003cem\u003eIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020)\u003c/em\u003e, Seattle, WA, USA, Jun. 2020, pp. 9726\u0026ndash;9735.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eK. M. He, X. Y. Zhang, S. Q. Ren, J. Sun, and Ieee, \u0026ldquo;Deep residual learning for image recognition,\u0026rdquo; in \u003cem\u003eProc.\u003c/em\u003e 2016 \u003cem\u003eIEEE Conference on Computer Vision and Pattern\u003c/em\u003e Recognition \u003cem\u003e(CVPR)\u003c/em\u003e, Las Vegas, NV, USA, Jun. 2016, pp. 770\u0026ndash;778.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eX. Huang et al., \u0026ldquo;Label-free 8-part leukocyte differential enabled by imaging flow cytometry of acoustic focusing and deep neural network of ResNet-18,\u0026rdquo; \u003cem\u003eBiomedical Instrumentation\u003c/em\u003e, vol. 1, no. 1, pp. 10004, Aug. 2025.\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":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"microsystems-and-nanoengineering","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"micronano","sideBox":"Learn more about [Microsystems \u0026 Nanoengineering](http://www.nature.com/micronano/)","snPcode":"41378","submissionUrl":"https://mts-micronano.nature.com/","title":"Microsystems \u0026 Nanoengineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Self-supervised learning, Imaging flow cytometry, Leukocyte differential, Strong generalization ","lastPublishedDoi":"10.21203/rs.3.rs-7956597/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7956597/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eImaging flow cytometry with supervised learning can realize high-accuracy leukocyte classification. However, since supervised learning relies on annotated cell images and this labeling process generates cell losses, current imaging flow cytometry with supervised learning cannot be used for leukocyte differential.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethodology\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study proposes a self-supervised contrastive learning framework pretrained on a non-annotated dataset comprising bright-field leukocyte images acquired via a custom imaging flow cytometry. The pretrained feature extractor, frozen during downstream tasks, employs lightweight linear or multilayer perceptron classifier heads to realize leukocyte differential in three independent annotated datasets: normal leukocyte subtypes (4-class), leukemia cells (2-class), and mixed normal/abnormal leukocytes (8-class).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResult\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe proposed method achieves classification accuracies of 96.16%, 96.68%, and 92.24% across three independent datasets, performing comparably to supervised baselines (with differences of less than 2%), therefore demonstrating high accuracy for leucocyte differentials. Furthermore, since the abnormal leukocytes in the latter two datasets were not included in the pre-training dataset, the corresponding F1-scores of 96.68% and 91.92% indicate the method’s strong generalization ability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe presented self-supervised learning framework enables high-accuracy and strong-generalization leukocyte differential using non-annotated images from imaging flow cytometry.\u003c/p\u003e","manuscriptTitle":"Generalizable Self-supervised Learning for Imaging Flow Cytometry on Multi-dataset Leukocyte Differential","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-19 11:49:29","doi":"10.21203/rs.3.rs-7956597/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"revise","date":"2025-12-30T00:52:48+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"This content is not available.","date":"2025-12-25T10:25:49+00:00","index":3,"fulltext":"This content is not available."},{"type":"editorInvitedReview","content":"This content is not available.","date":"2025-12-16T06:45:43+00:00","index":1,"fulltext":"This content is not available."},{"type":"reviewerAgreed","content":"This content is not available.","date":"2025-12-06T09:12:34+00:00","index":3,"fulltext":"This content is not available."},{"type":"reviewerAgreed","content":"This content is not available.","date":"2025-12-02T10:02:48+00:00","index":2,"fulltext":"This content is not available."},{"type":"reviewerAgreed","content":"This content is not available.","date":"2025-11-30T12:26:24+00:00","index":1,"fulltext":"This content is not available."},{"type":"reviewersInvited","content":"","date":"2025-11-07T07:53:59+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-28T04:03:48+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-25T15:37:21+00:00","index":"","fulltext":""},{"type":"submitted","content":"Microsystems \u0026 Nanoengineering","date":"2025-10-25T15:37:20+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"microsystems-and-nanoengineering","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"micronano","sideBox":"Learn more about [Microsystems \u0026 Nanoengineering](http://www.nature.com/micronano/)","snPcode":"41378","submissionUrl":"https://mts-micronano.nature.com/","title":"Microsystems \u0026 Nanoengineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5c47836c-1a88-40af-a1ac-0f1e793697af","owner":[],"postedDate":"November 19th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":57597830,"name":"Physical sciences/Engineering/Electrical and electronic engineering"},{"id":57597831,"name":"Physical sciences/Nanoscience and technology/Nanobiotechnology/Biosensors"}],"tags":[],"updatedAt":"2026-02-04T05:50:28+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-19 11:49:29","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7956597","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7956597","identity":"rs-7956597","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}