ABSTRACT
White blood cell differential (WBC‐Diff) scattergrams are used to screen for hematological abnormalities in potential hematopoietic malignancies and abnormal conditions. However, these potential indications can be easily missed. We applied deep learning models to develop a new and rapid workflow by learning the characteristics of WBC‐Diff scattergram images for early hematological abnormality screening. A total of 4297 WBC‐Diff scattergram images, including healthy controls, were included and divided into 19 categories. Eleven models encompassing both CNNs (ResNet18, ResNet50, VGG16, InceptionV3, DenseNet121, MobileNetV2, EfficientNet‐B0, and EfficientNet‐B1) and Vision Transformers (Swin‐Tiny, DeiT‐Base, and ConvNeXt‐Tiny) were trained and compared. Grad‐CAM interpretability analysis, hierarchical clustering, and multivariate analysis of variance (MANOVA) were used to characterize the learned feature space. Among all models, Swin‐Tiny achieved the highest accuracy (70.44%), while EfficientNet‐B0 was selected as the primary model due to its favorable balance of classification performance (67.90% accuracy) and computational efficiency (4.0 M parameters, 0.42 G floating‐point operations (GFLOPs), 2.8 × faster CPU inference than Swin‐Tiny). The 5‐run stability evaluation confirmed reproducible performance, and evaluation on an independent hold‐out dataset confirmed patient‐independent generalizability. As an upstream triage endpoint (normal vs. abnormal), the model achieved a sensitivity of 0.975 (test) and 0.984 (independent hold‐out set) with a false‐negative rate of 2.5% and 1.6%, respectively. At the screening level, only 2.2% of abnormal images were mis‐routed to “normal,” reflecting a sensitivity‐first, specificity‐as‐safeguard design. Grad‐CAM analysis revealed disease‐specific activation patterns that mapped to established cell population zones on the WBC‐Diff scattergram (e.g., myeloblast zone for M1 and neutrophil–basophil cluster for CML), providing support for the biological plausibility of model decisions, with significant differences among disease categories in the visual feature space (MANOVA, p < 0.01). This study demonstrated that a new automated hematology workflow incorporating EfficientNet‐B0 could provide early and rapid screening of potential hematological abnormalities, particularly suitable for deployment in resource‐constrained clinical settings.
Keywords: EfficientNet‐B0, hematological abnormalities, WBC‐Diff scattergrams
1. Introduction
The peripheral blood white blood cell (WBC) differential count is a time‐honored hematological laboratory test that provides some information and diagnostic clues even before peripheral smear examination [1, 2]. Automation generates numerical data and WBC scattergram graphic displays, such as WBC differentials (WBC‐Diff), WBC basophil (WBC/BASO), and immature myeloid information (IMI) channel scatter plots [2]. The WBC‐Diff graphic display patterns can provide clues regarding malaria, mononucleosis syndrome, and hematopoietic malignancies [3, 4, 5]. In the DIFF channel, cells are plotted according to side scatter light intensity (SS; x‐axis), reflecting internal cellular complexity and granularity, versus fluorescence intensity (FL; y‐axis), reflecting nucleic acid content. This two‐dimensional representation generates characteristic cell population clusters: lymphocytes occupy the low‐SS, moderate‐FL region; monocytes appear at moderate SS and FL; neutrophils and basophils cluster in the low‐FL, moderate‐SS zone; eosinophils are positioned at low FL with high SS; and immature granulocytes (IG) extend into the high‐SS region. Abnormal cell populations, including blasts, atypical lymphocytes, and plasma cells, typically appear in higher FL zones of the scattergram.
Hematopoietic malignancies are always characterized by abnormal, immature, or blast cell morphology in peripheral blood, which is detected by automated hematology analyzers as abnormal complete blood count (CBC) results and WBC scattergrams, indicating diseases such as acute myeloid leukemia (AML), chronic myelomonocytic leukemia (CMML), and acute lymphocytic leukemia (ALL) [6, 7, 8]. Although the diagnosis of these malignancies depends on molecular‐level analyses and gene expression [5, 9, 10], WBC graphic displays always provide useful information on disease discovery. Therefore, rapidly predicting hematopoietic diseases is important for patient diagnosis and treatment.
Many studies worldwide have investigated hematopoietic malignancy detection using automated hematology analyzers [11, 12, 13]. Abnormal particles may be observed on hematology analyzer WBC‐Diff scattergrams. Currently, automated hematology analyzers (e.g., Sysmex XN‐9000, XE‐5000) generate WBC‐Diff scattergrams, and previous studies have noted that these scattergrams may alert operators to abnormal samples, such as through detection of blasts, atypical lymphocytes, or immature granulocytes [11, 12, 13, 14]. However, this information does not always alert the operator, especially low numbers of abnormal cells (blasts, lymphoma cells, and dysplastic granulocytes) [14], and a relationship with any specific disease and these WBC graphic displays has not been reported. Therefore, strategies exclusively focusing on screening for abnormal particles, especially for hematopoietic malignancies, are needed beyond the current review rules.
Recently, convolutional neural networks (CNNs) have been applied to recognize cell subtypes in peripheral blood and bone marrow smears [15, 16, 17]. Studies have also used CNNs to count cells in equine bronchoalveolar lavage samples and other classification tasks [18, 19]. A CNN model was successfully applied for screening a special WBC‐Diff scattergram for acute promyelocytic leukemia (APL), which eliminated the potential pitfalls of manual observation [20]. Deep learning (DL) is a popular state‐of‐the‐art tool for image segmentation, which has been applied widely by researchers to develop accurate solutions to complex problems involving big image data [21], multi‐class diseases [22, 23], small datasets [24], class imbalance issues [25, 26], and the weak generalization ability of a single CNN trained with limited data [27]. All of these applications identified DL as a new method that could easily perceive minute changes in images and find abnormalities for disease detection.
In this study, we innovatively tried to map a special WBC‐Diff scattergram of hematopoietic malignancies and infectious mononucleosis (IM) to develop a deep learning‐based method for screening above conditions. This framework is challenged by multiple classes, small datasets, and disease imbalance issues. This was a multimodal approach consisting of IM, AML without maturation (M1), AML with maturation (M2), APL (M3), acute myelomonocytic leukemia (M4), acute monoblastic/monocytic leukemia (M5), ALL, CLL, chronic myelogenous leukemia (CML), CMML, immune thrombocytopenia (ITP), myelodysplastic syndrome with excess blasts (MDS)‐EB, MDS with multilineage dysplasia (MDS‐MLD), MDS with ring sideroblasts (MDS‐RS), MDS with lineage dysplasia (MDS‐SLD), B‐cell non‐Hodgkin lymphoma (B‐NHL), T‐cell non‐Hodgkin lymphoma (T‐NHL), multiple myeloma (MM), and healthy controls (HCs). To the best of our knowledge, this is the first DL multinomial multi‐classification focusing on screening for hematopoietic malignancies and IM using WBC‐Diff scattergram images. Pretrained state‐of‐the‐art DL approaches were applied, including eight CNN architectures (ResNet18, ResNet50, VGG16, InceptionV3, DenseNet121, MobileNetV2, EfficientNet‐B0, and EfficientNet‐B1) and three Vision Transformer architectures (Swin‐Tiny, DeiT‐Base, and ConvNeXt‐Tiny) to comprehensively evaluate both convolutional and attention‐based paradigms for this classification task. These approaches offered possible solutions and advanced architectures for multi‐class classification. Therefore, the major goal of this study was to improve WBC‐Diff scattergram image analysis, which is often bottlenecked by limited and non‐independent data arising from multiple images per patient.
2. Methods
2.1. Study Design and Participants
The CNN multinomial multi‐classification development and validation study was conducted using data from the First Affiliated Hospital of Chongqing databases. All relevant cases between January 1, 2016, and December 31, 2024, were recruited. Inclusion criteria: complete clinical records were available for all patients. Hematology patients (e.g., acute myeloid leukemia and myelodysplastic syndromes) were diagnosed by bone marrow cytomorphology, flow cytometry, and/or genetic typing. Infectious disease patients (e.g., infectious mononucleosis) were diagnosed by EBV antibody testing. ITP (immune thrombocytopenia) patients were diagnosed based on the clinical opinion of the attending physician, combined with clinical manifestations of thrombocytopenia, exclusion of other secondary thrombocytopenias (e.g., autoimmune diseases, infection, and drugs), and reference to auxiliary examinations such as complete blood count and bone marrow examination (e.g., megakaryocyte count and morphology). Only the blood routine test (with corresponding WBC‐Diff scattergram) obtained before any chemotherapy at the time of initial diagnosis was included for patients with hematological malignancies; for ITP patients, only samples obtained before any immunosuppressive therapy at the time of initial diagnosis were included. All included blood routine samples were reviewed by manual microscopy. Exclusion criteria: blood routine reports obtained after radiotherapy, chemotherapy, blood transfusion, targeted drug therapy, or immunosuppressive therapy. Hematology patients without a clear diagnosis or with concurrent severe non‐hematological diseases. Healthy control WBC‐Diff scattergram images were retrospectively retrieved from the hospital's laboratory information system (LIS) for individuals who had undergone routine health checkups during the study period. A healthy control was defined as an individual with a complete blood count (CBC) within the normal reference range (hemoglobin, white blood cell count, platelet count, and differential count), no documented history of hematological disorders, chronic inflammatory, autoimmune, or infectious diseases, and no record of medications affecting hematopoiesis within the past 3 months. A normal WBC‐Diff scattergram was defined as a scattergram from a healthy control meeting these criteria, with no hematological flags generated by the Sysmex automated hematology analyzer (e.g., “Blast?”, “Abnormal Lympho?”, and “Immature Granulocyte?”). This approach ensures that the scattergram represents a typical normal leukocyte population distribution, aligning with the clinical use of analyzer flags as objective indicators of sample abnormality.
A total of 4297 WBC‐Diff scattergram images (samples), including healthy controls, were included; per‐sample instrument flag data (e.g., blast flag and atypical lymph flag) were not recorded during the data collection process; because the de‐identified images do not carry an explicit patient identifier, all counts are reported at the image/sample level. These patients were diagnosed with IM, AML, ALL, CLL, CML, CMML, MM, ITP, MDS, and lymphoma. AML was divided into M1, M2, M3, M4, and M5. Lymphoma was divided into B‐NHL and T‐NHL. MDS was divided into categories, namely MDS‐SLD, MDS‐RS, MDS‐MLD, and MDS‐EB. For model‐training labels, AML subtypes were recorded using the French–American–British (FAB) categories (M1–M5) and MDS subtypes using the subclassification in routine institutional use during 2016–2024; these labels reflect this historical local diagnostic grouping rather than strict WHO 5th‐edition genetic/lineage entities.
AML, ALL, CLL, CML, CMML, MM, MDS, lymphoma, and disease categories were diagnosed according to the World Health Organization Fifth Edition Classification Project [28]. HCs comprise healthy donors undergoing a routine checkup who had normal WBC‐Diff scattergrams. A total of 4297 WBC‐Diff scattergram images, acquired predominantly on Sysmex analyzers (XN‐9000, XN‐1000, XE‐2100, and XE‐5000) with a small number (n = 3, all infectious mononucleosis cases) on a Mindray BC‐660 analyzer, were correlated and divided into 19 subfolders according to the above diagnoses. In addition, Grad‐CAM29 was used to generate pseudo‐color heatmaps highlighting key features influencing model classification, which were overlaid on the original image for intuitive visualization of the model's attention regions. Differences among images were analyzed by multivariate analysis of variance (MANOVA) using Python.
2.2. Image Dataset Preprocessing and Augmentation
In this study, all images were resized to a resolution of 224 × 224 pixels using bilinear interpolation and normalized using ImageNet channel means ([0.485, 0.456, and 0.406]) and standard deviations ([0.229, 0.224, and 0.225]) to stabilize gradient‐based optimization. Generally, if the number of initial images is limited, CNNs cannot be trained as designed; thus, the data augmentation method is used to address the overfitting problem [26]. The specific offline augmentation operations included random scaling (scaling factors ranging from 0.9 to 1.1) and adjustments to brightness, contrast, and saturation. Horizontal flips, vertical flips, and rotations were deliberately excluded because WBC‐Diff scatter plots are orientation‐sensitive; the spatial distribution of cell population clusters on the SS–FL axes carries inherent diagnostic significance. During model training, additional online augmentation was applied via random cropping (resize to 256 pixels followed by random crop to 224 pixels) and color jitter (brightness, contrast, saturation ± 0.1; hue ±0.02). The number of images for each class is shown in Table S1. Augmentation was applied exclusively to the training set after dataset splitting; the validation, test, and independent hold‐out sets contained only original, non‐augmented images, and all reported validation, test, and independent metrics were computed on original images only. Augmented files were named “<class>_<index>_aug<NN>” to ensure traceability, and an audit confirmed that no augmented image appeared in any evaluation set (File S1, Section S1.3). The augmentation parameters were applied with class‐specific expansion factors: 2× for majority classes and up to 6× for underrepresented classes, as detailed in Section 2.4.
2.3. Dataset Splits for Model Training, Validation, and Evaluation
We divided randomly the datasets into a training set (60%), validation set (20%), and test set (20%) using stratified splitting to maintain class proportions across subsets. Because the de‐identified scattergram filenames do not carry an explicit patient identifier, data splitting was initially performed at the image level, as filenames did not carry explicit patient identifiers. A post hoc acquisition‐metadata audit was conducted to assess potential patient‐level data leakage. This audit found no duplicate same‐sample/acquisition across the training, validation, and test sets. However, we acknowledge that this audit method cannot exclude the possibility that two different samples from the same patient, collected on different dates, were assigned to different partitions. The number of distinct patients and the maximum number of images per patient for the rare classes (CMML, MDS‐SLD, MDS‐MLD, MDS‐RS) are reported in the Section 3 (see Table S6). Therefore, patient‐level leakage remains a primary limitation. We consider patient‐level leakage, particularly for MDS‐RS and CMML, a primary limitation rather than a resolved issue. As detailed in the Methods, the high average number of images per patient in these classes (e.g., 5.0 for MDS‐RS) makes patient‐level splitting essential yet infeasible in our retrospective design. This leakage likely explains the sharp recall drop observed for MDS‐RS (1.00 on the test set vs. 0.25 on the hold‐out set). Future studies should implement patient‐level data splitting from the outset. To assess potential patient‐level leakage, we performed a post hoc audit using MD5 content hashing against the raw archive and parsing of acquisition metadata (analyzer model, acquisition date, and daily sample sequence number). Sample keys were reconstructed for 96.6% of files, and no duplicate same‐sample/acquisition key from the training set appeared in the validation or test set (Table 1; File S1). However, this audit cannot exclude the possibility that two different blood samples from the same patient, collected on different dates, were assigned to different partitions. Critically, retrospective manual tracing revealed that for rare classes with very few patients, such as MDS‐RS (2 patients contributing 10 training images, average 5 images/patient) and CMML (3 patients contributing 20 training images, average 6.7 images/patient), the high average number of images per patient strongly suggests that patient‐level leakage is virtually unavoidable. This limitation is discussed in detail in Section 4 (see also Supplementary for the number of distinct patients per class across all subsets). This approach ensured that the distribution of different categories remained consistent across all subsets, as displayed in Table S2. Additionally, to evaluate patient‐independent generalizability, an independent hold‐out dataset comprising 500 WBC‐Diff scattergram images was held out entirely from training, validation, and testing. These images were acquired during 2025 and shared no patients or images with the development cohort, as confirmed by manual tracing of patient identifiers (File S1). This cohort encompassed all 19 categories, with sample sizes ranging from 4 (MDS‐RS) to 93 (B‐NHL), reflecting the natural clinical prevalence distribution. WBC‐Diff scattergram images were acquired from routine hematology analyzer workflows with standardized acquisition protocols. No formal quality filtering was applied to exclude poor‐quality samples; images were included based on clinical diagnosis and analyzer availability, as the standardized acquisition protocols at our institution ensure consistent image quality. For the development cohort, the training set comprised 2568 original images that were expanded offline to 7956 images for model fitting, whereas the validation (n = 863), test (n = 866), and independent hold‐out (n = 500) sets comprised original (non‐augmented) images only; the complete flow of images from enrollment through splitting, augmentation, and evaluation is summarized in Figure S1.
TABLE 1.
Summary of the data‐provenance and split audit (File S1).
| Verification check | Outcome |
|---|---|
| Augmented images in val/test/independent set | 0/0/0 |
| Identical original‐image stem across splits | 0/4297 |
| Reconstructed sample keys crossing train–val–test | 0 |
| Sample keys crossing test–independent | 1 (false positive; two different patients) |
| Files with acquisition metadata recovered | 9513/9849 (96.6%) |
| Image/metadata overlap, development‐2025 vs. independent | 0/0 |
2.4. Model Training and Validation
We evaluated 11 models encompassing both CNN architectures and modern Vision Transformers. The CNN models included ResNet18, ResNet50, VGG16, InceptionV3, DenseNet121, EfficientNet‐B0 [24], and EfficientNet‐B1. To provide contemporary baselines and address the evolving landscape of computer vision, three representative Transformer‐based architectures were added: Swin‐Tiny (hierarchical shifted windows), DeiT‐Base (data‐efficient image Transformer), and ConvNeXt‐Tiny (modernized convolutional design). All models were initialized with ImageNet‐pretrained weights and fine‐tuned with a two‐stage learning rate strategy (backbone at 0.1× the head learning rate).
The fully connected layers of all models were replaced with new classification layers containing 19 output units to adapt to the task requirements. The training was conducted using the AdamW optimizer (initial learning rate = 5 × 10−5, weight decay = 1 × 10−3) and a cosine annealing learning rate scheduler incorporating a 5‐epoch linear warm‐up phase. Focal Loss (γ = 2.0) with class‐specific weights was employed as the loss function, combined with label smoothing (ε = 0.05) to reduce overconfidence. Each model was trained for up to 100 epochs, with early stopping (patience = 15 epochs) triggered when the validation performance showed no further improvement. The model checkpoint with the best validation performance was retained for final evaluation. To address class imbalance evident in, we employed three complementary strategies: (1) stratified splitting to preserve class proportions across training, validation, and test sets; (2) offline data augmentation with variable rates (2× for majority classes, up to 6× for underrepresented classes); and (3) a BalancedBatchSampler during training to ensure each mini‐batch contained representative samples from minority classes.
2.5. Model Evaluation Metrics
Model performance was evaluated on the held‐out test set (20%) after training was completed. During evaluation, the network operated in inference mode, generating predicted class labels based on the maximum output probability of the SoftMax layer. For each model, we computed overall accuracy, macro‐averaged precision, recall, and F1‐score across all 19 classes, as well as per‐class metrics including class‐specific accuracy, precision, recall, F1‐score, and specificity. A confusion matrix was generated by comparing model predictions to ground‐truth labels and visualized as a heat‐map for intuitive assessment of classification patterns. A normalized (row‐wise) confusion matrix was also produced to directly display per‐class recall. To assess model stability, EfficientNet‐B0 was trained five times with different random seeds, and the mean and standard deviation of all metrics were reported. The trained EfficientNet‐B0 model was further applied to the independent hold‐out dataset (Section 2.3) without any retraining or fine‐tuning, evaluating the same set of metrics to assess patient‐independent generalizability. In addition to the 19‐class metrics, we evaluated three clinically oriented endpoints by collapsing the multiclass predictions: normal vs. abnormal; malignant/suspicious vs. non‐malignant/reactive; and smear‐review‐needed vs. not‐needed. For each endpoint, we computed sensitivity, specificity, positive and negative predictive values, false‐negative rate, and the area under the ROC curve. Bootstrap resampling (2000 iterations) was used to obtain 95% confidence intervals for all main, per‐class, and binary‐endpoint metrics, and for device‐stratified analyses. Because the model is intended as an upstream screening/triage tool, the clinically primary metrics are sensitivity (recall) and the false‐negative rate, particularly at the binary screening level (normal vs. abnormal), since a missed abnormal case carries substantially greater clinical cost than a false alarm; specificity is treated as a safeguard that constrains the false‐positive rate and hence the downstream smear‐review workload. Performance is therefore interpreted within a sensitivity‐first, specificity‐as‐safeguard framework, with multiclass metrics characterizing finer subtype discrimination.
Gradient‐weighted class activation mapping (Grad‐CAM) heatmaps were generated for all 19 categories using the final convolutional layer (features[−1]) of EfficientNet‐B0 as the target layer. For each test sample, the class activation map (CAM) was computed with respect to its true class label. To visualize the Grad‐CAM feature space, per‐sample CAM vectors (7 × 7 = 49 dimensions) were projected into two dimensions using t‐distributed stochastic neighbor embedding (t‐SNE; perplexity = 30, 2000 iterations). Dot sizes in the resulting scatter plot were scaled proportionally to the maximum activation intensity of each sample. Additionally, to characterize the spatial distribution of model attention, each class‐average CAM was down‐sampled to a 3 × 3 grid, where each cell represents the mean activation intensity in the corresponding spatial region (top‐left to bottom‐right). These 3 × 3 activation maps were individually normalized and gamma‐corrected (γ = 0.55) to enhance visual contrast. The top‐5 best‐performing and top‐5 worst‐performing classes (ranked by recall) were further analyzed by displaying Grad‐CAM overlays for correctly classified and misclassified samples. Because Grad‐CAM maps were computed with respect to the true class label, they reflect the regions relevant to the correct class and therefore do not fully explain the model's actual decision in misclassified cases; predicted‐label maps would be required for that purpose. We accordingly interpret these maps as evidence of interpretability and biological plausibility rather than biological validation.
2.6. Statistical Analysis and Software
All models were implemented and trained using the PyTorch deep learning framework with Python. The ImgAug30 library was used for image augmentation preprocessing. Hierarchical clustering was applied to the aforementioned distance matrix derived from the Grad‐CAM results. Ward's linkage method was employed to minimize intra‐cluster variance during cluster merging.
To assess the statistical significance of differences in hematological cell type in the visual feature space, we conducted MANOVA.
3. Results
3.1. Evaluation of Training Convergence and Generalization Across All Models
To address data transparency and directly assess the potential for patient‐level leakage, we performed a retrospective analysis of patient identifiers for all 19 classes. The number of distinct patients per class, along with the maximum and average number of images per patient, across the training, validation, test, and independent hold‐out sets, is reported in Table S6.
Based on the acquisition‐metadata audit, no duplicate same‐sample/acquisition was found across partitions. However, this audit cannot exclude the possibility that two different samples from the same patient, collected on different dates, were assigned to different partitions. Critically, our retrospective manual tracing revealed that for the rarest classes, such as MDS‐RS (2 patients contributing 10 training images, average 5 images/patient) and CMML (3 patients contributing 20 training images, average 6.7 images/patient), the high average number of images per patient strongly suggests that patient‐level leakage is virtually unavoidable. This is a primary limitation of this study, and its potential impact on the performance of rare classes (e.g., the MDS‐RS recall drop) is discussed in detail in Section 4. The training and validation loss curves for all 11 models are presented in Figure 1. Figure 1A shows that all models achieved decreasing training loss over epochs, with EfficientNet‐B0 and DeiT‐Base converging to the lowest training loss values (< 0.1) after approximately 50 epochs. However, Figure 1B reveals important differences in generalization behavior. Swin‐Tiny exhibited the lowest and most stable validation loss throughout training (63 epochs), indicating superior generalization capacity. In contrast, VGG16 exhibited continuously rising validation loss from early epochs, consistent with severe overfitting likely attributable to its large parameter count (134.3 M). EfficientNet‐B0 demonstrated a characteristic pattern: rapid initial convergence followed by a widening train‐validation loss gap after approximately 20 epochs, suggesting progressive overfitting despite its compact architecture (4.0 M parameters). The early stopping mechanism preserved the best‐performing checkpoint, mitigating the impact of this overfitting tendency on final test performance. Notably, the three Transformer‐based architectures (Swin‐Tiny, DeiT‐Base, ConvNeXt‐Tiny, shown as dashed lines) generally maintained more stable validation loss trajectories compared to most CNN models, suggesting that attention mechanisms may offer inherent regularization benefits for this classification task. The corresponding training and validation accuracy curves are shown in Figure 1C,D, where EfficientNet‐B0 reached approximately 100% training accuracy but plateaued at ~65% validation accuracy, consistent with the overfitting pattern observed in the loss curves.
FIGURE 1.
Training dynamics and multimodel performance comparison. (A) Training loss curves over epochs for all 11 models (eight CNN architectures shown with solid lines; three Transformer‐based models shown with dashed lines). EfficientNet‐B0 (red, thick line) is highlighted as the primary model. (B) Corresponding validation loss curves. The widening gap between training and validation loss for EfficientNet‐B0 after ~20 epochs reflects a moderate overfitting tendency, though early stopping preserved the optimal checkpoint. (C) Training accuracy curves over epochs for all 11 models. (D) Validation accuracy curves over epochs. EfficientNet‐B0 reaches ~100% training accuracy but plateaus at ~65% validation accuracy, consistent with the overfitting pattern observed in the loss curves. (E) Grouped bar chart comparing Accuracy, Precision, Recall, and F1‐Score for all 11 models, ranked by test accuracy. Transformer models are highlighted with shaded backgrounds; EfficientNet‐B0 is marked with a red highlight. (F) Scatter plot of test accuracy versus total parameters (millions), showing the accuracy‐efficiency tradeoff. EfficientNet‐B0 achieves the highest accuracy‐to‐parameter ratio among the top‐three models.
Among the 11 models compared (Table 2), Swin‐Tiny achieved the highest test accuracy (70.44%) and F1‐score (67.90%), followed by DeiT‐Base (68.94%, 66.20%) and EfficientNet‐B0 (67.90%, 65.62%). ConvNeXt‐Tiny (67.67%), EfficientNet‐B1 (66.28%), ResNet18 (66.05%), ResNet50 (64.67%), MobileNetV2 (64.55%), and DenseNet121 (64.20%) showed comparable mid‐range performance, while InceptionV3 (62.59%) and VGG16 (59.82%) exhibited relatively lower accuracy. Importantly, when considering computational efficiency, EfficientNet‐B0 required only 0.42 G floating‐point operations (GFLOPs) and 23.0 ms per image on CPU, compared to 4.51 GFLOPs and 63.4 ms for Swin‐Tiny—representing a 10.7× reduction in computational cost and 2.8× faster inference speed. EfficientNet‐B0 achieved the second‐highest accuracy‐to‐parameter ratio (16.84% per million parameters) among all models, outperformed only by MobileNetV2 (28.71), which had substantially lower absolute accuracy. EfficientNet‐B0's recall (65.77%) ranked second overall, suggesting strong detection capability across disease categories—a critical attribute for clinical screening applications. The comparative performance across all four metrics is visualized in Figure 1E. Figure 1F further illustrates the accuracy‐versus‐parameter trade‐off, highlighting EfficientNet‐B0's position as the most parameter‐efficient model among the top performers (4.0 M parameters, 35.9 MB GPU memory).
TABLE 2.
Performance and computational efficiency comparison of 11 deep learning models for WBC‐Diff scattergram multi‐class classification.
| Model | Params (M) | GFLOPs | CPU inference (ms) | GPU memory (MB) | Accuracy | Precision (macro) | Recall (macro) | F1‐score (macro) |
|---|---|---|---|---|---|---|---|---|
| Swin‐Tiny | 27.5 | 4.51 | 63.4 | 137.2 | 0.7044 | 0.7178 | 0.6614 | 0.6790 |
| DeiT‐Base | 85.8 | 16.87 | 59.8 | 360.9 | 0.6894 | 0.6867 | 0.6540 | 0.6620 |
| EfficientNet‐B0 | 4.0 | 0.42 | 23.0 | 35.9 | 0.6790 | 0.6725 | 0.6577 | 0.6562 |
| ConvNeXt‐Tiny | 27.8 | 4.47 | 28.1 | 134.3 | 0.6767 | 0.7043 | 0.6085 | 0.6247 |
| EfficientNet‐B1 | 6.5 | 0.62 | 31.1 | 46.1 | 0.6628 | 0.6854 | 0.5995 | 0.6135 |
| ResNet18 | 11.2 | 1.83 | 14.8 | 64.6 | 0.6605 | 0.6510 | 0.6179 | 0.6237 |
| ResNet50 | 23.5 | 4.14 | 35.0 | 116.8 | 0.6467 | 0.6541 | 0.5957 | 0.6130 |
| MobileNetV2 | 2.2 | 0.33 | 17.6 | 28.9 | 0.6455 | 0.6246 | 0.5979 | 0.6038 |
| DenseNet121 | 7.0 | 2.91 | 53.7 | 48.4 | 0.6420 | 0.6388 | 0.5989 | 0.6032 |
| InceptionV3 | 21.8 | 2.86 | 42.0 | 107.9 | 0.6259 | 0.6343 | 0.6086 | 0.5970 |
| VGG16 | 134.3 | 15.47 | 60.0 | 585.2 | 0.5982 | 0.6077 | 0.5769 | 0.5758 |
Note: Models are ranked by test accuracy in descending order. EfficientNet‐B0 (bold) was selected as the primary model for clinical deployment. GFLOPs (giga floating‐point operations per second) quantifies computational complexity; CPU inference time was measured with single‐image batch size on an Intel CPU; GPU memory was measured as peak allocation on an NVIDIA RTX 4060 Ti.
Rather than performing exhaustive hyperparameter grid search, we adopted a pre‐validated training configuration (learning rate = 5 × 10−5, batch size = 16, AdamW with cosine annealing and warm‐up) informed by established best practices for transfer learning with EfficientNet architectures. This decision was motivated by two considerations: (1) the advanced training strategies already employed (Focal Loss, BalancedBatchSampler, layered learning rates, and early stopping) substantially reduce the sensitivity to individual hyperparameter choices; and (2) the primary performance bottleneck lies in the inherent difficulty of 19‐class classification with limited minority‐class samples rather than suboptimal hyperparameters. To verify this and assess model stability, EfficientNet‐B0 was trained five times with different random seeds (Table 3). The mean test accuracy was 0.6728 ± 0.0080 and mean F1‐score was 0.6459 ± 0.0109, with a narrow coefficient of variation (CV = 1.2% for accuracy), confirming that performance was reproducible and not critically dependent on initialization or stochastic training dynamics. Furthermore, evaluation on an independent hold‐out dataset confirmed the model's patient‐independent generalizability beyond the original training cohort, as detailed in the next section.
TABLE 3.
Five‐run stability evaluation of EfficientNet‐B0 with different random seeds.
| Seed | Accuracy | Precision | Recall | F1 (macro) | Specificity |
|---|---|---|---|---|---|
| 42 | 0.6790 | 0.6722 | 0.6580 | 0.6562 | 0.9824 |
| 123 | 0.6663 | 0.6563 | 0.6396 | 0.6360 | 0.9823 |
| 456 | 0.6836 | 0.6781 | 0.6665 | 0.6593 | 0.9824 |
| 789 | 0.6617 | 0.6538 | 0.6339 | 0.6315 | 0.9823 |
| 1024 | 0.6732 | 0.6665 | 0.6503 | 0.6465 | 0.9824 |
| Mean ± SD | 0.6728 ± 0.0080 | 0.6654 ± 0.0092 | 0.6497 ± 0.0118 | 0.6459 ± 0.0109 | 0.9824 ± 0.0001 |
Note: EfficientNet‐B0 was trained five times with different random seeds to assess reproducibility. All other hyperparameters were held constant. The mean ± SD row summarizes performance variability across runs.
3.2. Overall Performance Evaluation of the EfficientNet‐B0 Model for Multi‐Class Classification Tasks
WBC‐Diff scattergram images were fed into EfficientNet‐B0 model to predict the potential occurrence of each hematological abnormality class for validation. Figure 2A shows the confusion matrices obtained on the test dataset, and Figure 2B presents the corresponding normalized confusion matrix illustrating per‐class recall. The per‐class performance is further visualized in Figure 2C, where classes are ranked by recall. Seven classes achieved F1‐scores above 0.80: CML (0.92), MDS‐RS (0.89), M3 (0.87), CLL (0.87), IM (0.86), M1 (0.82), and NORMAL (0.81), indicating robust classification for these well‐represented disease categories. CMML achieved an F1‐score of 0.77. Moderate performance (F1 0.55–0.67) was observed for ALL, MDS‐MLD, M5, MM, ITP, B‐NHL, and M2. The lowest recall was observed for MDS‐SLD (0.17), M4 (0.38), and T‐NHL (0.33), likely reflecting their limited sample sizes and overlapping scattergram morphology with other disease categories. Specificity exceeded 0.95 for all 19 classes, yet high one‐vs‐rest specificity coexisted with low recall for several rare classes in this imbalanced setting (per‐class bootstrap 95% CIs are reported in Table 4, which juxtaposes test‐set and independent hold‐out metrics, and in File S1). Recall confidence intervals were very wide for low‐prevalence classes (MDS‐MLD [0.00, 1.00]; MDS‐SLD [0.00, 0.57]; M4 [0.13, 0.62]) but narrow for well‐represented classes (M3 [0.75, 0.88]; CML [0.83, 0.98]). Importantly, the per‐class performance on the test set was comparable to that on the validation set (test accuracy 0.6790 vs. validation accuracy 0.6698; test macro‐F1 0.6562 vs. validation macro‐F1 0.6255), with no evidence of overfitting (Table 4; Table S3).
FIGURE 2.
Detailed classification performance of EfficientNet‐B0 on the test dataset. (A) Confusion matrix showing absolute sample counts for the 19‐class classification task. Diagonal cells indicate correctly classified samples; color intensity reflects sample frequency. (B) Normalized confusion matrix, where each row is scaled by the total number of true samples per class, directly visualizing per‐class recall. (C) Per‐class performance bar chart showing Accuracy, Precision, Recall, and F1‐Score for each of the 19 disease classes, sorted by Recall in descending order. The green dashed line indicates the mean Recall across all classes (65.8%). High‐recall classes (e.g., MDS‐RS, IM, and CML) demonstrate robust detection, while low‐recall classes (e.g., MDS‐SLD, T‐NHL, and M4) reflect limited sample sizes and morphological overlap.
TABLE 4.
Per‐class performance comparison between the test set and independent hold‐out set for EfficientNet‐B0, with bootstrap 95% confidence intervals.
| Class | Test_Precision [95% CI] | Test_Recall [95% CI] | Test_F1 [95% CI] | Independent hold‐out_Precision [95% CI] | Independent hold‐out_Recall [95% CI] | Independent hold‐out_F1 [95% CI] | Recall_Δ |
|---|---|---|---|---|---|---|---|
| ALL | 0.6622 [0.549, 0.765] | 0.6712 [0.560, 0.776] | 0.6667 [0.566, 0.748] | 0.6562 [0.486, 0.815] | 0.5526 [0.395, 0.719] | 0.6000 [0.453, 0.727] | −0.1186 |
| B‐NHL | 0.5978 [0.500, 0.696] | 0.5093 [0.414, 0.607] | 0.5500 [0.466, 0.629] | 0.6447 [0.536, 0.744] | 0.5269 [0.427, 0.625] | 0.5799 [0.489, 0.663] | +0.0176 |
| CLL | 0.8421 [0.652, 1.000] | 0.8889 [0.722, 1.000] | 0.8649 [0.727, 0.970] | 0.7500 [0.524, 0.944] | 1.0000 [1.000, 1.000] | 0.8571 [0.688, 0.971] | +0.1111 |
| CML | 0.9318 [0.849, 1.000] | 0.9111 [0.825, 0.979] | 0.9213 [0.857, 0.974] | 1.0000 [1.000, 1.000] | 0.9706 [0.905, 1.000] | 0.9851 [0.950, 1.000] | +0.0595 |
| CMML | 0.8333 [0.500, 1.000] | 0.7143 [0.333, 1.000] | 0.7692 [0.400, 1.000] | 1.0000 [1.000, 1.000] | 0.8333 [0.500, 1.000] | 0.9091 [0.667, 1.000] | +0.1190 |
| IM | 0.8158 [0.688, 0.935] | 0.9118 [0.800, 1.000] | 0.8611 [0.765, 0.939] | 0.7222 [0.500, 0.929] | 0.8667 [0.667, 1.000] | 0.7879 [0.606, 0.927] | −0.0451 |
| ITP | 0.6400 [0.444, 0.818] | 0.5517 [0.357, 0.735] | 0.5926 [0.417, 0.735] | 0.5500 [0.400, 0.703] | 0.6667 [0.500, 0.824] | 0.6027 [0.464, 0.722] | +0.1150 |
| M1 | 0.7500 [0.455, 1.000] | 0.9000 [0.667, 1.000] | 0.8182 [0.588, 0.960] | 0.9091 [0.700, 1.000] | 1.0000 [1.000, 1.000] | 0.9524 [0.824, 1.000] | +0.1000 |
| M2 | 0.5294 [0.418, 0.656] | 0.5373 [0.419, 0.656] | 0.5333 [0.434, 0.633] | 0.5476 [0.395, 0.697] | 0.5227 [0.368, 0.674] | 0.5349 [0.400, 0.654] | −0.0146 |
| M3 | 0.9344 [0.888, 0.975] | 0.8143 [0.750, 0.875] | 0.8702 [0.824, 0.910] | 0.8889 [0.787, 0.974] | 0.8000 [0.692, 0.902] | 0.8421 [0.759, 0.917] | −0.0143 |
| M4 | 0.3158 [0.111, 0.533] | 0.3750 [0.133, 0.619] | 0.3429 [0.129, 0.529] | 0.1818 [0.000, 0.444] | 0.2222 [0.000, 0.539] | 0.2000 [0.000, 0.444] | −0.1528 |
| M5 | 0.6154 [0.510, 0.702] | 0.6882 [0.589, 0.776] | 0.6497 [0.561, 0.720] | 0.4615 [0.305, 0.613] | 0.5806 [0.394, 0.759] | 0.5143 [0.351, 0.641] | −0.1076 |
| MDS‐EB | 0.4800 [0.273, 0.684] | 0.4138 [0.238, 0.609] | 0.4444 [0.263, 0.603] | 0.6667 [0.455, 0.875] | 0.6364 [0.417, 0.840] | 0.6512 [0.457, 0.800] | +0.2226 |
| MDS‐MLD | 1.0000 [0.000, 1.000] | 0.5000 [0.000, 1.000] | 0.6667 [0.000, 1.000] | 0.5000 [0.000, 1.000] | 0.4286 [0.000, 0.833] | 0.4615 [0.000, 0.769] | −0.0714 |
| MDS‐RS | 0.8000 [0.331, 1.000] | 1.0000 [1.000, 1.000] | 0.8889 [0.498, 1.000] | 0.3333 [0.000, 1.000] | 0.2500 [0.000, 1.000] | 0.2857 [0.000, 0.667] | −0.7500 |
| MDS‐SLD | 0.3333 [0.000, 1.000] | 0.1667 [0.000, 0.571] | 0.2222 [0.000, 0.600] | 0.6667 [0.000, 1.000] | 0.4000 [0.000, 1.000] | 0.5000 [0.000, 0.889] | +0.2333 |
| MM | 0.5814 [0.478, 0.685] | 0.7353 [0.632, 0.841] | 0.6494 [0.561, 0.730] | 0.5000 [0.364, 0.643] | 0.5476 [0.394, 0.700] | 0.5227 [0.388, 0.641] | −0.1877 |
| NORMAL | 0.7561 [0.659, 0.846] | 0.8732 [0.787, 0.943] | 0.8105 [0.736, 0.874] | 0.5556 [0.320, 0.783] | 0.7143 [0.462, 0.938] | 0.6250 [0.400, 0.800] | −0.1589 |
| T‐NHL | 0.3590 [0.212, 0.514] | 0.3333 [0.192, 0.478] | 0.3457 [0.208, 0.473] | 0.3143 [0.161, 0.480] | 0.3548 [0.188, 0.531] | 0.3333 [0.176, 0.471] | +0.0215 |
Note: For each of the 19 classes, precision, recall, and F1‐score are reported for both the test set and independent hold‐out set. The independent hold‐out support column shows the number of samples per class in the independent hold‐out set. Recall_Δ shows the difference in recall (independent hold‐out − test).
When the multiclass predictions were collapsed into clinically oriented endpoints, the model performed strongly as an upstream triage tool. For normal vs. abnormal, sensitivity was 0.975 (95% CI 0.962–0.985) on the test set and 0.984 on the independent hold‐out set, with false‐negative rates of 2.5% and 1.6% and AUCs of 0.990 and 0.971, respectively. For malignant/suspicious vs. non‐malignant/reactive, sensitivity was 0.958 (test) and 0.934 (independent), and for smear‐review‐needed vs. not‐needed, 0.965 and 0.945. Full operating characteristics (sensitivity, specificity, PPV, NPV, FNR, and AUC, each with bootstrap 95% confidence intervals) are reported in Table 5. Because positive and negative predictive values depend strongly on disease prevalence, we plotted prevalence‐adjusted PPV and NPV for the three clinically oriented endpoints across a 1%–99% abnormal‐prevalence range using the independent hold‐out sensitivity and specificity (Figure 3). For the normal‐vs‐abnormal endpoint, PPV rose steeply with prevalence (e.g., 0.15 at 5% and 0.78 at 50%), whereas NPV remained high (> 0.97) at the low prevalence typical of screening. Notably, the low NPV observed on the independent set (0.556) is a direct consequence of its enriched abnormal prevalence (97.2%): at this operating point the same sensitivity and specificity yield an NPV of ≈0.56, whereas at realistic screening prevalence the projected NPV approaches 1.0. The predictive values therefore reflect the specific case mix and should not be extrapolated across settings; sensitivity and specificity are the transferable operating characteristics.
TABLE 5.
Clinically oriented binary endpoints derived from the EfficientNet‐B0 multiclass predictions, with bootstrap 95% confidence intervals (2000 iterations).
| Endpoint | Set | Sensitivity | Specificity | PPV | NPV | FNR | AUC |
|---|---|---|---|---|---|---|---|
| Normal vs. Abnormal | Test | 0.975 [0.963, 0.985] | 0.873 [0.787, 0.943] | 0.989 [0.980, 0.995] | 0.756 [0.659, 0.846] | 0.025 [0.015, 0.037] | 0.990 [0.983, 0.995] |
| Normal vs. Abnormal | Independent | 0.984 [0.971, 0.994] | 0.714 [0.455, 0.938] | 0.992 [0.983, 0.998] | 0.556 [0.312, 0.778] | 0.016 [0.006, 0.029] | 0.971 [0.933, 0.993] |
| Malignant vs. Reactive | Test | 0.958 [0.943, 0.971] | 0.851 [0.790, 0.912] | 0.972 [0.959, 0.985] | 0.786 [0.720, 0.851] | 0.042 [0.029, 0.057] | 0.976 [0.962, 0.987] |
| Malignant vs. Reactive | Independent | 0.934 [0.909, 0.956] | 0.758 [0.652, 0.862] | 0.965 [0.947, 0.981] | 0.618 [0.507, 0.729] | 0.066 [0.044, 0.091] | 0.950 [0.927, 0.970] |
| Smear‐needed vs. Not | Test | 0.965 [0.951, 0.977] | 0.800 [0.719, 0.876] | 0.974 [0.962, 0.985] | 0.748 [0.664, 0.830] | 0.035 [0.023, 0.049] | 0.971 [0.952, 0.984] |
| Smear‐needed vs. Not | Independent | 0.945 [0.923, 0.965] | 0.702 [0.566, 0.828] | 0.968 [0.951, 0.984] | 0.569 [0.436, 0.694] | 0.055 [0.035, 0.077] | 0.951 [0.928, 0.971] |
FIGURE 3.
Prevalence‐adjusted predictive values for three clinical endpoints (EfficientNet‐B0, independent hold‐out se/sp). (A) Positive predictive value (PPV) and (B) negative predictive value (NPV) as a function of abnormal (disease) prevalence (1%–99%) for the normal‐vs‐abnormal, malignant/suspicious‐vs‐reactive, and smear‐review‐needed‐vs‐not‐needed endpoints, each computed from its own independent hold‐out sensitivity and specificity. The shaded band denotes the 95% CI of the normal‐vs‐abnormal endpoint, propagated from the bootstrap CIs of sensitivity and specificity; the gray area marks the low‐prevalence range typical of screening. The dashed line marks the abnormal prevalence of the present enriched cohort (97.2%), at which NPV ≈0.56 for the normal‐vs‐abnormal endpoint—explaining the low NPV reported in Table 5—whereas at realistic screening prevalence the projected NPV approaches 1.0. PPV/NPV are prevalence‐dependent, whereas sensitivity and specificity are the transferable metrics.
To directly quantify screening safety, we distinguished true screening misses (abnormal cases predicted as normal, which would not be referred for review) from confusion among abnormal subtypes (which are referred regardless). Of 1281 abnormal images, only 28 (2.2%) were predicted as normal (2.5% on the test set, 1.6% on the independent hold‐out set), corresponding to a screening sensitivity of 97.8% (Table 5; Figures 2A and 6A). In contrast, 33.2% of abnormal images were assigned to a different abnormal subtype; because every abnormal prediction triggers smear review, such inter‐subtype confusion does not constitute a screening miss. Consequently, the moderate multiclass recall (macro recall 0.658) does not translate into missed abnormal cases. The corresponding false‐positive rate—normal images flagged as abnormal—was 12.7% on the test set, representing an acceptable additional smear‐review burden.
FIGURE 6.
Independent hold‐out validation performance of the EfficientNet‐B0 Model on an independent dataset (n = 500). (A) Confusion matrix showing absolute sample counts for the 19‐class classification task on the independent hold‐out dataset. Diagonal cells indicate correctly classified samples; color intensity reflects sample frequency. (B) Normalized (row‐wise) confusion matrix on the independent hold‐out set, scaling each row by the number of true samples per class to directly display per‐class recall. (C) Dumbbell chart comparing per‐class recall between the original test set (blue dots) and the independent hold‐out set (red dots) for all 19 disease classes, sorted by independent hold‐out recall in ascending order. Sample sizes (n) for each class in the independent hold‐out set are annotated on the right. (C) Scatter plot of test‐set recall (x‐axis) versus independent hold‐out recall (y‐axis) for each disease class, with dot sizes proportional to the independent hold‐out sample count. The dashed diagonal line represents perfect concordance (y = x). Pearson and Spearman correlation coefficients are annotated. Classes near the diagonal indicate consistent performance in the independent hold‐out cohort.
3.3. Grad‐CAM Interpretability and Activation Region Analysis
Individual Grad‐CAM heatmaps for all 19 classes (Figure 4A) revealed that the model consistently focused on distinct scattergram regions corresponding to disease‐specific cell population distributions, with red areas indicating regions of highest discriminative importance. Hierarchical clustering of class‐average Grad‐CAM vectors (Figure 4B) identified three major clusters: Cluster 1 (orange) included a mix of myeloid neoplasms (M1, MDS‐RS, CML, and M3), lymphoid malignancies (ALL and MM), a reactive condition (ITP), and normal controls, suggesting that the model's activation patterns in the lymphocyte‐dominant and monocyte/blast zones overlap across these classes. Cluster 2 (green) comprised CLL and IM, both characterized by prominent lymphocyte activation patterns. Cluster 3 (red) grouped M4, M5, MDS subtypes, B‐NHL, T‐NHL, CMML, and M2, reflecting shared activation patterns in the granulocyte and monocyte regions of the scattergram. This mixed clustering indicates that the Grad‐CAM activation patterns are driven by the spatial distribution of cell populations on the scattergram rather than by lineage alone, consistent with the known cell population topology of the WBC‐Diff channel. To further characterize the Grad‐CAM feature space, t‐SNE projection of per‐sample activation vectors (Figure 4C, left panel) demonstrated partial but meaningful separation among disease classes, with dot sizes reflecting activation intensity. The 3 × 3 spatial activation maps (Figure 4C, right panel) quantified the regional distribution of model attention for each class. Notably, M1 showed strong central activation, CML and MM exhibited bottom‐shifted activation patterns, and MDS‐RS displayed a distinctive lower‐right activation profile. These spatial activation patterns are consistent with the established cell population topology of WBC‐Diff scattergrams. Specifically, the central activation in M1 maps to the monocyte/blast zone (moderate FL and SS) where myeloblasts are characteristically located; the bottom‐shifted pattern in CML aligns with the neutrophil–basophil cluster (low FL and moderate SS), consistent with the hallmark granulocytic hyperplasia of this disease; and the lower‐right activation in MDS‐RS corresponds to the high‐SS region, potentially reflecting dysplastic granulocytes with altered granularity [29]. Moreover, the hierarchical clustering separation of lymphoid diseases (CLL, MM, B‐NHL, T‐NHL, NORMAL) from myeloid neoplasms mirrors the spatial segregation of lymphocyte and myeloid cell populations on the scattergram, indicating that the model captured biologically plausible, lineage‐specific activation patterns rather than arbitrary image features. Significant differences were observed among disease categories in the visual feature space (MANOVA, p < 0.01; Table S4; pairwise post hoc comparisons among classes are detailed in Table S5), confirming the discriminative value of the learned features.
FIGURE 4.
Grad‐CAM Interpretability Analysis of EfficientNet‐B0 for WBC‐Diff Scattergram Classification. (A) Grad‐CAM heatmaps for all 19 disease classes. Each panel shows a representative correctly classified test sample overlaid with the Grad‐CAM activation map. Red regions indicate areas of highest discriminative importance for the model's classification decision. (B) Hierarchical clustering dendrogram of class‐average Grad‐CAM activation vectors. Ward's linkage was applied to Euclidean distances between 49‐dimensional (7 × 7) mean activation vectors. The dendrogram is colored by the three major clusters identified: Orange for Cluster 1 (M1, MDS‐RS, CML, M3, NORMAL, ALL, ITP, and MM), green for Cluster 2 (CLL and IM), and red for Cluster 3 (M4, M5, and MDS subtypes [MDS‐MLD, MDS‐SLD, and MDS‐EB], B‐NHL, T‐NHL, CMML, and M2). The blue horizontal line indicates the global highest level branch cut. The clustering reveals a mix of myeloid, lymphoid, and reactive conditions, reflecting overlapping activation patterns in the scattergram's cell population zones (e.g., lymphocyte‐dominant and monocyte/blast zones) rather than a clean lineage‐based separation. This pattern is consistent with the spatial distribution of cell populations on the WBC‐Diff scattergram. Disease classes with similar spatial activation patterns cluster together, revealing clinically meaningful groupings. (C) High‐activation region analysis. Left panel: T‐SNE projection of per‐sample Grad‐CAM activation vectors (7 × 7 = 49 dimensions), colored by disease class. Dot sizes are proportional to maximum activation intensity, with larger dots indicating stronger model confidence. Right panel: Class‐average 3 × 3 spatial activation maps showing the normalized mean Grad‐CAM intensity per spatial region. White grid lines delineate the nine subregions (T/M/B × L/C/R). White cross markers indicate peak activation positions. The color scale ranges from low activation (deep blue) to high activation (red), with the custom academic colormap matching the scatter plot color scheme.
3.4. Representative Case Analysis of Top‐Performing and Worst‐Performing Disease Classes
To provide concrete visual evidence of model behavior, we examined the Grad‐CAM activation patterns for the top‐5 best‐performing and top‐5 worst‐performing disease classes by recall. The top‐5 best‐performing classes (Figure 5A: MDS‐RS, IM, CML, M1, and CLL) showed highly consistent Grad‐CAM patterns for correctly classified samples, with focused activation in disease‐specific scattergram regions. Notably, MDS‐RS achieved 100% recall with no misclassified samples. These consistent patterns suggest that these disease classes possess distinctive, learnable scattergram features. In contrast, the top‐5 worst‐performing classes (Figure 5B: MDS‐SLD, T‐NHL, M4, MDS‐EB, and MDS‐MLD) exhibited diffuse and variable activation patterns in misclassified samples. The predicted labels revealed common confusion pairs (e.g., MDS‐SLD → MDS‐EB, T‐NHL → MM, M4 → M2), suggesting that classification errors arise from morphological overlap between disease subtypes and may be ameliorated by increasing training sample sizes for minority classes.
FIGURE 5.
Grad‐CAM Analysis of Top‐5 Best‐Performing and Top‐5 Worst‐Performing Disease Classes (EfficientNet‐B0). (A) Top‐5 best‐performing classes by recall (MDS‐RS, IM, CML, M1, and CLL). Left three columns show Grad‐CAM overlays for correctly classified test samples (green border); right three columns show misclassified samples (red border) with the predicted class label. MDS‐RS achieved 100% recall with no misclassified samples. The consistently focused activation patterns in correctly classified samples suggest that these disease classes have distinctive, learnable scattergram features. (B) Top‐5 worst‐performing classes by recall (MDS‐SLD, T‐NHL, M4, MDS‐EB, and MDS‐MLD). The same layout as 4 A. Misclassified samples show diffuse, unfocused activation patterns, and the predicted labels reveal common confusion pairs (e.g., MDS‐SLD → MDS‐EB, T‐NHL → MM, and M4 → M2). These patterns suggest that classification errors arise from morphological overlap between disease subtypes and may be ameliorated by increasing training sample sizes for minority classes.
3.5. Independent Hold‐Out Validation of the EfficientNet‐B0 Model
To evaluate the patient‐independent generalizability of the EfficientNet‐B0 model, an independent hold‐out dataset of 500 WBC‐Diff scattergram images, sharing no images or patient‐level samples with the development cohort, was evaluated (File S1). The confusion matrix (Figure 6A) illustrates the model's classification performance across all 19 categories on this independent hold‐out cohort. The overall accuracy was 62.4%, with macro‐averaged precision of 62.4%, recall of 62.5%, and F1‐score of 61.8% (Table 6), closely matching the test‐set performance (accuracy 67.9%, F1 65.6%) and confirming that the model did not suffer from significant performance drift. Figure 6B presents a dumbbell chart comparing per‐class recall between the test set and the independent hold‐out set (detailed per‐class metrics for both sets, together with the per‐class recall differences, are reported in Table 4). High‐recall classes on the test set (e.g., CML, CLL, IM, and M1) maintained strong performance in the independent hold‐out set (CML: 0.97, CLL: 1.00, M1: 1.00). For MDS‐RS, the independent hold‐out sample size was only 4; while the model achieved perfect recall, this estimate is highly uncertain and should be interpreted with caution. The performance may be inflated due to the small number of cases. Conversely, classes with limited independent hold‐out samples (e.g., MDS‐RS: n = 4, MDS‐SLD: n = 5) exhibited greater variability. To quantify the concordance of per‐class performance, Figure 6C plots test‐set recall against independent hold‐out recall for each class, with dot sizes proportional to the independent hold‐out sample count. The Pearson correlation coefficient confirmed a significant positive association between the two evaluation settings, indicating that the model's relative strengths and weaknesses were preserved in the independent hold‐out cohort. Device‐stratified analysis showed comparable performance: XN‐9000 accuracy 0.661 (95% CI 0.609–0.712) and XE‐5000 accuracy 0.553 (95% CI 0.482–0.629). The bootstrap‐based difference in accuracy was 0.108 (95% CI: −0.012 to 0.228), indicating no statistically significant difference between the two analyzers. The slightly lower point estimate for XE‐5000 likely reflects its smaller sample size and rare‐class instability rather than a systematic instrument bias; the lower XE‐5000 point estimate is consistent with its smaller sample size and the instability of rare‐class metrics (File S1, Section S1.5). A normalized (row‐wise) confusion matrix for this cohort is provided in Figure 6B, directly displaying per‐class recall: high‐recall classes such as CLL and M1 (recall = 1.00) were preserved, whereas rare or morphologically overlapping classes (e.g., M4, MDS‐RS, and T‐NHL) showed lower recall, consistent with the test‐set pattern. Grad‐CAM heatmaps computed on the independent hold‐out set (Figure 7, the independent‐set counterpart of Figure 4A) showed that, for well‐recognized classes such as CLL, CML, and M1, activation remained spatially consistent with the development set, whereas for ITP, M4, MM, NORMAL, and T‐NHL, the attention maps were more diffuse and shifted relative to Figure 4A, paralleling the lower recall of these classes.
TABLE 6.
Summary of EfficientNet‐B0 performance on the test set and independent hold‐out set.
| Metric | Test set | Independent hold‐out | Δ (Independent − Test) |
|---|---|---|---|
| Accuracy | 0.6790 | 0.6240 | −0.0550 |
| Precision (macro) | 0.6725 | 0.6236 | −0.0489 |
| Recall (macro) | 0.6577 | 0.6249 | −0.0327 |
| F1‐Score (macro) | 0.6562 | 0.6182 | −0.0380 |
| Sample size | 866 | 500 | — |
Note: Macro‐averaged metrics are reported. The independent hold‐out dataset (n = 500) was collected from a different time period than the training data. Δ represents the difference between independent and test‐set metrics.
FIGURE 7.
Grad‐CAM heatmaps on the independent hold‐out set (EfficientNet‐B0). Each panel shows a representative correctly classified image from the independent hold‐out cohort for one of the 19 disease classes, overlaid with its Grad‐CAM activation map computed at the final convolutional layer (features[−1]). Warm colors indicate regions of highest discriminative importance. For well‐recognized classes (e.g., CLL, CML, and M1) activation matched the development set, whereas for ITP, M4, MM, NORMAL, and T‐NHL it was more diffuse and shifted relative to Figure 4A (see main text).
3.6. Workflow Incorporating the EfficientNet‐B0 Model for Early Hematological Abnormality Screening
Using the metrics for the EfficientNet‐B0 multi‐class model, we constructed an innovative workflow of WBC‐Diff results analysis with enhanced sensitivity for early hematological abnormality screening that incorporated the above model framework. The cascaded workflow steps are shown in Figure 8. First, blood samples were analyzed using a hematology analyzer that had WBC‐Diff scattergram outputs. Next, the scattergrams were evaluated using the EfficientNet‐B0 model to screen for suspicious abnormal mapping. Third, a positive predictive WBC‐Diff scattergram result was obtained to ensure screening accuracy. Therefore, a more modern pipeline for WBC‐Diff scattergram analysis was constructed by introducing the EfficientNet‐B0 model, where rapid hematological abnormality screening could be configured upstream to increase the efficiency of blood smear examinations using machine‐learning‐based, morphological classifications. This pipeline is designed for upstream triage screening prior to morphological review; prospective multicenter validation, device‐specific robustness testing, calibration analysis, and assessment of smear‐review workload and missed abnormal cases remain priorities for future work. Accordingly, Figure 8 should be regarded as a proposed decision‐support/triage framework rather than a validated clinical pathway; its real‐world performance—particularly the predictive values that govern triage decisions—requires prospective validation under the substantially lower abnormal prevalence of routine screening settings before clinical deployment.
FIGURE 8.
Proposed clinical workflow incorporating the EfficientNet‐B0 model for early hematological abnormality screening. This workflow is presented as a conceptual decision‐support/triage framework, not a validated clinical deployment tool. Prospective, multicenter, and calibration studies are warranted before clinical deployment. Blood samples are analyzed by an automated hematology analyzer generating WBC‐Diff scattergram outputs. These scattergrams are then evaluated by the EfficientNet‐B0 model to screen for suspicious abnormal patterns, with positive predictions triggering further morphological review.
4. Discussion
We present a new analysis protocol for the detection and characterization of hematological diseases using WBC scattergram images of automated hematology analyzers linked with DL models (CNN and Vision Transformers) that could analyze these WBC scattergram images to distinguish clinical conditions and identify novel associations between images and clinical diagnoses. This new protocol can screen WBC‐Diff scattergrams and provide the earliest information for suspicious hematological diseases, even before a morphological review. This protocol serves as a comprehensive supplement, not limited to identifying suspicious APL cases [20] but also provides hematological disease screening markers that facilitate early detection prior to initiating morphological reexamination procedures. Importantly, our approach could be generalized to other centers to apply similar principles to WBC scattergram images in automated hematology analyzers. For instance, conventional alarm rules on hematology analyzers have reported false‐negative rates of less than 5% for clinically significant abnormalities [30], whereas our model achieved a false‐negative rate of only 1.6%–2.5% in the normal/abnormal triage endpoint. Additionally, prior deep‐learning studies focused on single diseases (e.g., APL) [20, 31], while our framework covers 19 categories with per‐class specificity exceeding 0.95. This demonstrates a substantial improvement in screening breadth and accuracy compared with current practice.
The EfficientNet‐B0 model proposed in this study is not intended to replace the built‐in flagging functions of automated hematology analyzers, but to serve as an upstream triage complement: blood samples are first analyzed by the hematology analyzer to generate WBC‐Diff scattergrams; the scattergrams are then evaluated by the model to screen for suspicious abnormal patterns; positive predictions trigger subsequent morphological review. By learning global leukocyte population distributions from the scattergram image itself, the model can capture indirect signals (e.g., lower‐right activation for MDS‐RS and bottom‐shifted activation for MM) that may be overlooked by conventional cell‐count‐based flagging algorithms. Thus, the model and the existing flags are complementary: the flags focus on specific cell‐type alerts, while the model provides a more comprehensive, image‐level screening with multi‐class output.
We acknowledge that the present study did not perform a direct head‐to‐head comparison between the model and instrument flags, because per‐sample flag data (e.g., blast flag and atypical lymph flag) were not recorded during the retrospective data collection. Nevertheless, the model's sensitivity in the normal/abnormal binary endpoint was 0.975 on the test set and 0.984 on the independent hold‐out set, with false‐negative rates of only 2.5% and 1.6%, respectively; per‐class specificity exceeded 0.95 for all 19 classes; and no myeloid malignancy was missed (all eight false negatives in the test set were lymphoid diseases). These results indirectly suggest the model's potential added value relative to existing flags. Future studies should prospectively collect instrument flag data and perform a quantitative comparison with the model's predictions, particularly in the low‐burden regime where abnormal cells are scarce, to rigorously demonstrate the incremental benefit of the proposed approach.
This study did not perform a dedicated subgroup analysis of samples with low abnormal cell burden (e.g., blast percentage < 5%), which is a limitation. However, the model's very low false‐negative rates on the overall test set and independent hold‐out set (2.5% and 1.6%), and the fact that no myeloid malignancy was missed, indirectly suggest that the model may still have good screening ability in the low‐burden setting. Furthermore, Grad‐CAM analysis revealed that the model focuses on disease‐specific scattergram regions (e.g., lower‐right activation for MDS‐RS, bottom‐shifted activation for MM). Even when abnormal cells are scarce, the relative density changes in the affected cell population may still be detectable by the model. Future studies should specifically collect samples with predefined low abnormal cell percentages (e.g., < 5% blasts or < 1% abnormal lymphocytes) for independent validation to quantify the model's performance in this critical scenario and to clarify its incremental value over existing flags.
Similar to the technological limitations of manual morphology identification, the ability to screen abnormal WBC scattergram images varies with different laboratory technician experience levels. Although some previous studies have summarized the characteristics of WBC scattergram images in conditions such as malaria, mononucleosis syndrome, and hematological malignancies [3, 5, 29], the new CNN EfficientNet‐B0 model introduced in our study could screen 19 kinds of hematological diseases with high efficiency, accuracy, and precision. We simplified the protocol and improved the efficiency before the morphological review.
Importantly, our Grad‐CAM interpretability analysis provides the first systematic evidence linking DL model attention to the established cell population biology of WBC‐Diff scattergrams. Previous studies have described WBC scattergram abnormalities in individual diseases such as malaria, mononucleosis, and specific hematological malignancies [3, 5, 29], but a comprehensive mapping between model‐derived activation patterns and the underlying cell population topology has not been reported. Our results demonstrate that the model's disease‐specific activation regions correspond to the expected locations of pathologically relevant cell populations on the DIFF channel—for example, myeloblasts in the moderate‐FL/SS zone for M1, expanded granulocytes in the low‐FL zone for CML, and the lineage‐level separation of lymphoid versus myeloid diseases in hierarchical clustering in Figure 4B—consistent with the biological plausibility of the model's classification decisions. Notably, the model appears to discriminate among disease classes based on characteristic changes in cell population density and cluster morphology within specific scattergram regions, rather than relying solely on absolute spatial coordinates—consistent with the underlying hematological pathology, where different diseases produce distinct patterns of cell population expansion, depletion, or morphological alteration on the WBC‐Diff scattergram. For MM, the bottom‐shifted activation may reflect secondary changes in the overall leukocyte composition rather than direct plasma cell detection, as circulating plasma cells are relatively uncommon in peripheral blood except in advanced disease. Similarly, the lower‐right activation for MDS‐RS may be attributable to dysplastic granulocytes with altered granularity rather than ring sideroblasts themselves, as erythroid precursors are not directly represented on the WBC‐Diff channel. Future studies integrating Grad‐CAM [32] analysis with concurrent multiparameter scattergram data (e.g., WBC/BASO and IMI channels) may further elucidate the biological basis of these disease‐specific activation signatures.
Among 11 models including both CNNs and Transformers, Swin‐Tiny achieved the highest overall accuracy (70.44%). However, EfficientNet‐B0 was selected as the primary model due to its favorable balance of classification performance (67.90% accuracy, third‐highest overall) and computational efficiency. Specifically, EfficientNet‐B0 required only 0.42 GFLOPs and achieved a CPU inference time of 23.0 ms per image (≈43 images/s), compared to 4.51 GFLOPs and 63.4 ms for Swin‐Tiny. This 2.8× speed advantage on CPU is particularly relevant for deployment in primary hospitals and resource‐constrained clinical laboratories where dedicated GPU hardware may not be available, enabling real‐time screening integrated with routine hematology analyzer workflows. The training curves (Figure 1B) revealed a characteristic overfitting pattern for EfficientNet‐B0; nevertheless, the early stopping mechanism preserved the best‐performing checkpoint, and the 5‐run stability evaluation (Table 3) confirmed reproducible test performance. The inclusion of Transformer architectures confirmed that attention‐based models can capture scattergram spatial patterns effectively, with Swin‐Tiny maintaining the most stable validation loss trajectory. However, their substantially larger parameter counts (7× to 21× that of EfficientNet‐B0) and higher computational demands may limit deployment in settings without dedicated computing infrastructure. Future work may explore knowledge distillation from Swin‐Tiny to EfficientNet‐B0 or stronger regularization strategies to further improve classification accuracy while preserving the lightweight architecture advantage.
Crucially, the evaluation on an independent hold‐out dataset of 500 samples (patient‐independent and single‐center) demonstrated that the EfficientNet‐B0 model maintained comparable performance to the original test set (Table 4), with no evidence of significant performance drift. The per‐class recall comparison (Figure 6B) revealed that high‐performing classes such as CML, CLL, and M1 retained excellent recall in the independent hold‐out cohort, while the overall ranking of class‐level performance remained largely consistent, as confirmed by the significant positive correlation between test‐set and independent hold‐out recall (Figure 6C). Notably, some classes showed improved performance on the independent hold‐out set (e.g., MDS‐EB and CMML), possibly reflecting richer morphological diversity in the independent hold‐out cohort for these categories. Conversely, classes with very small independent hold‐out sample sizes exhibited unstable metrics, consistent with limited representation of rare disease subtypes in this cohort. These independent hold‐out results strengthen the clinical applicability of the model by confirming that learned scattergram features are not artifacts of a specific data collection period.
The balance between high sensitivity and high specificity observed here reflects a deliberate sensitivity‐first, specificity‐as‐safeguard design rationale appropriate for triage. In a pre‐smear‐review setting, the two error types are highly asymmetric: a false negative may delay the detection of a hematological malignancy, whereas a false positive merely incurs an additional smear review—a low‐cost, reversible action—so prioritizing sensitivity is clinically justified. Two further points reinforce this. First, the high discrimination of the screening endpoint (AUC 0.99 on the test set and 0.97 on the independent hold‐out set) indicates that the decision threshold can be shifted toward even higher sensitivity, should a lower false‐negative rate be required, at a modest cost in specificity. Second, the per‐class specificity exceeding 0.95 acts as the safeguard that keeps the overall false‐positive load—and hence the smear‐review burden—within a clinically manageable range. Notably, the few screening misses occurred predominantly in lymphoid/plasma‐cell entities (B‐NHL, MM, and T‐NHL), whose circulating cells share the lymphocyte region of the DIFF scattergram—consistent with plasma cells being terminally differentiated (activated) B lymphocytes—and may therefore overlap with the normal lymphocyte distribution; this biological plausibility, rather than a generic model failure, suggests these subtypes as priorities for future feature enrichment.
Justification of included classes. We justify including ITP, MM, and B−/T‐NHL based on three lines of evidence. First, these diseases induce indirect leukocyte population changes detectable on scattergrams (e.g., bottom‐shifted activation for MM, lower‐right activation for MDS‐RS). Second, although lymphoid recall was lower, per‐class specificity exceeded 0.95 for all 19 classes, and all eight false negatives in the test set were lymphoid (no myeloid entity missed). This trade‐off is acceptable for a triage tool that prioritizes myeloid malignancy screening. Third, including a broad disease spectrum enhances clinical utility as a non‐specific abnormality screener, with mandatory secondary review for lymphoid cases. The independent hold‐out validation confirmed stable performance, with high recall for myeloid classes (CML 0.97, CLL 1.00, and M1 1.00). Minority classes (MDS‐SLD and T‐NHL) remain limited by sample size, which we acknowledge as a limitation. Furthermore, our study has a significant limitation regarding patient‐level data leakage. The original data split was performed at the image level. While a post hoc acquisition‐metadata audit found no duplicate samples across partitions 16, this audit cannot rule out the possibility that two different samples from the same patient, collected on different days, were assigned to different sets. This is a particularly critical concern for rare classes with very few samples, as the high average number of images per patient in these classes (e.g., MDS‐RS: 2 patients contributing 10 training images, average 5 images/patient; CMML: 3 patients contributing 20 training images, average 6.7 images/patient) makes patient‐level leakage virtually unavoidable. The sharp recall drop observed for MDS‐RS (1.00 on the test set vs. 0.25 on the hold‐out set) is most parsimoniously explained by this combination of extremely small sample size and unavoidable patient‐level leakage. We therefore consider this a primary limitation, not a resolved issue. Future studies should implement patient‐level data splitting from the outset to provide a more rigorous estimate of model generalizability. Future studies should implement patient‐level data splitting from the outset, ideally by stratifying patients before acquisition, to ensure the independence of training and validation cohorts. We have now reported the number of distinct patients per class in Section 3 to increase transparency.
In the proposed workflow, model predictions are combined with subsequent blood smear analysis, in which great advances have also been made by machine learning in the morphological analysis of blood cells [33, 34, 35]. The CNN models proved effective and reliable for tailoring therapeutical strategies to help clinicians in everyday clinical practice [36, 37]. We emphasize that this pipeline is proposed as a potential decision‐support/triage framework rather than a validated clinical deployment tool; prospective, multicenter, and calibration studies are warranted before clinical deployment. It is also worth noting that the current 19‐class model could be simplified to a direct binary classification model (normal vs. abnormal) for the triage task, which would likely improve performance on rare classes and reduce model complexity. We consider this a promising direction for future research.
Several limitations remain. As with other AI classification models, performance in minority classes depends on adequate per‐class sample representation [37]. Our findings indicate that EfficientNet‐B0, as a lightweight model, leverages a simple yet effective compound coefficient to achieve uniform scaling across depth, width, and resolution when extending CNNs. Currently, this neural network model holds great development potential in areas such as inferring genetic causal relationships, auxiliary diagnosis of infectious and chronic diseases, and cancer prognosis predictions [38, 39, 40, 41]. Additionally, no formal quality filtering was applied to the input images. While the standardized acquisition protocols at our institution ensured reasonable image quality, future work should explore automated quality assessment to further improve robustness. The lower recall observed in minority classes (e.g., MDS‐MLD and MDS‐SLD) likely reflects residual effects of class imbalance despite our mitigation strategies. Although the evaluation on an independent hold‐out dataset provided encouraging evidence of generalizability, all data were still collected from a single institution using four Sysmex analyzer models (XN‐9000, XN‐1000, XE‐2100, and XE‐5000) and, for a small number of images (n = 3), a Mindray BC‐660 analyzer. Multi‐center validation studies across different instruments and clinical settings are warranted to further establish the robustness and transferability of the model. Consistent with this, the classes showing the largest attention drift between the development and independent sets on Grad‐CAM (ITP, MM, and T‐NHL, and to a lesser extent M4 and NORMAL) are largely those for which a specific, reproducible leukocyte‐differential signature is not biologically expected: ITP is a platelet disorder, whereas MM and B−/T‐NHL are primarily marrow‐ or tissue‐based with inconstant peripheral‐blood expression. The concurrent instability of Grad‐CAM localization and per‐class recall in these entities is consistent with the absence of a stable, transferable scattergram feature rather than a modeling artifact alone, and indicates that WBC‐Diff‐based classification is most reliable for entities with a genuine differential signature, such as myeloid neoplasms.
Furthermore, Grad‐CAM interpretability analysis was performed exclusively on EfficientNet‐B0 because this technique is specifically designed for convolutional neural network architectures and requires gradient information from convolutional feature maps. Interpreting Transformer‐based models such as Swin‐Tiny requires alternative visualization approaches (e.g., attention rollout or class attention maps), which may yield complementary insights. Cross‐architecture interpretability comparison represents a promising direction for future research. In addition, because the present Grad‐CAM maps were generated using true class labels, future work incorporating predicted‐label maps would better characterize the model's decisions in misclassified cases.
A promising direction for future work is to train a dedicated binary classification model (normal vs. abnormal) for the triage task. Such a model would avoid the confusions between adjacent pathological subtypes that dominate the multi‐class errors, and would be particularly beneficial for improving the performance on rare disease classes. We plan to compare the performance of a direct binary classifier with the current multi‐class model in a future study. This approach would also simplify the model architecture and reduce computational requirements, making it more suitable for deployment in primary care settings.
5. Conclusion
Our work provides a new protocol to prompt early hematological abnormality screening using EfficientNet‐B0, which can eliminate the potential subjectivity, intervisibility, inadvisability, and irreproducibility of manual observations. The model demonstrated robust performance across 19 categories, with 5‐run stability evaluation confirming reproducible results and evaluation on an independent hold‐out dataset verifying its patient‐independent generalizability beyond the original training cohort. While the 19‐class model provides comprehensive diagnostic information, a dedicated binary classification model for normal/abnormal triage represents a simpler and potentially more accurate alternative for future clinical deployment. This would be a valuable direction for future work. Grad‐CAM interpretability analysis revealed that the model learned disease‐specific scattergram features consistent with known pathophysiological patterns. This protocol can help identify potentially abnormal results needing further examination for clinical diagnosis, especially in primary hospitals. With a screening sensitivity of 97.8% (false‐negative rate 2.2%), the workflow prioritizes detection of abnormal cases while maintaining specificity sufficient for practical deployment.
In addition, future studies should collect instrument flag data to perform head‐to‐head comparisons between the model and existing flagging functions, and to conduct dedicated validation in the low‐burden regime, in order to fully establish the clinical incremental value of this model.
Author Contributions
Haixia Wang: conceptualization, investigation, methodology, validation, visualization, writing – review and editing, project administration, funding acquisition, supervision, resources, formal analysis, data curation. Xuan Luo: data curation, writing – original draft, investigation, formal analysis. Te Chen: formal analysis, software, visualization, writing – review and editing, validation, methodology, project administration, data curation. Mingzhu Long: data curation, formal analysis, writing – original draft. Lixia Zhang: methodology, validation, investigation.
Funding
Research reported in this publication was supported by Chongqing medical scientific research project (Joint project of Chongqing Health Commission and Science and Technology Bureau) No. 2022MSXM120. The content is solely the responsibility of the authors and does not necessarily represent the official views of the Chongqing medical scientific research project.
Ethics Statement
All included participants signed an informed consent form. The study protocol was in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board of the First Affiliated Hospital of Chongqing Medical University. Research data are stored in an institutional repository. The secondary use of study data will be considered upon a reasonable request and approval of the Institutional Review Board. The data are not publicly available because of privacy or ethical restrictions.
Conflicts of Interest
The authors declare no conflicts of interest.
Supporting information
Figure S1: STARD‐style flow of WBC‐Diff scattergram images. The development cohort (n = 4297; 2016–2024) was split at the image level into training (n = 2568), validation (n = 863), and test (n = 866) sets (stratified; 60/20/20; random_state = 42). Offline augmentation was applied to the training set only, expanding it to 7956 images. An independent hold‐out cohort (n = 500; 2025), disjoint from the development cohort at both the image and patient level, was evaluated without any retraining or fine‐tuning. Validation, test, and independent hold‐out sets comprised original (non‐augmented) images only. A content‐hash (MD5) plus acquisition‐metadata audit confirmed no image‐ or patient‐level overlap across splits (File S1).
Table S1: The account of initial, expand quantity of augmentation, and total number in each class.
Table S2: The count of each class for dataset split model training, validation, and evaluation.
Table S3: Overfit detection: Comparison of validation set and test set performance for EfficientNet‐B0 multi‐class classification.
Table S4: Post hoc pairwise comparisons using Tukey's HSD test on PCA components derived from EfficientNet‐B0 Grad‐CAM Activation Vectors.
Table S5: Post hoc pairwise comparisons using Tukey's HSD test on PCA components derived from EfficientNet‐B0 Grad‐CAM activation vectors.
Table S6: Patient‐level data audit for the four rare hematological classes.
File S1: Data provenance and split audit report.
Acknowledgements
We thank LetPub (www.letpub.com.cn) for its linguistic assistance during the preparation of this manuscript.
Contributor Information
Te Chen, Email: chente@cqmu.edu.cn.
Haixia Wang, Email: czwhx@163.com.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author reasonable request. The data are not publicly available because of privacy or ethical restrictions.
References
- 1. Sandhaus L. M. and Meyer P., “How Useful Are CBC and Reticulocyte Reports to Clinicians?,” American Journal of Clinical Pathology 118, no. 5 (2002): 787–793, 10.1309/cqgg-hy0u-lrkl-glmp. [DOI] [PubMed] [Google Scholar]
- 2. Sireci A., Schlaberg R., and Kratz A., “A Method for Optimizing and Validating Institution‐Specific Flagging Criteria for Automated Cell Counters,” Archives of Pathology & Laboratory Medicine 134, no. 10 (2010): 1528–1533, 10.5858/2009-0531-oa.1. [DOI] [PubMed] [Google Scholar]
- 3. Sale S., Carone A. E., Fumi M., Pancione Y., and Rocco V., “Detection of Apoptotic Lymphocytes Through Sysmex XN‐1000 as a Diagnostic Marker for Mononucleosis Syndrome,” Journal of Clinical Laboratory Analysis 30, no. 5 (2016): 779–793, 10.1002/jcla.21938. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Post A., Kaboré B., Reuling I. J., et al., “The XN‐30 Hematology Analyzer for Rapid Sensitive Detection of Malaria: A Diagnostic Accuracy Study,” BMC Medicine 17, no. 1 (2019): 103, 10.1186/s12916-019-1334-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Gupta M., Chauhan K., Singhvi T., Kumari M., and Grover R. K., “Useful Information Provided by Graphic Displays of Automated Cell Counter in Hematological Malignancies,” Journal of Clinical Laboratory Analysis 32, no. 5 (2018): e22392, 10.1002/jcla.22392. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Pyzer A. R., Stroopinsky D., Rajabi H., et al., “MUC1‐Mediated Induction of Myeloid‐Derived Suppressor Cells in Patients With Acute Myeloid Leukemia,” Blood 129, no. 13 (2017): 1791–1801, 10.1182/blood-2016-07-730614. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Park D., Chang J., Kahng J., et al., “Development of a Novel Flow Cytometry‐Based System for White Blood Cell Differential Counts: 10‐Color LeukoDiff,” Annals of Laboratory Medicine 39, no. 2 (2019): 141–149, 10.3343/alm.2019.39.2.141. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Prada‐Arismendy J., Arroyave J. C., and Röthlisberger S., “Molecular Biomarkers in Acute Myeloid Leukemia,” Blood Reviews 31, no. 1 (2017): 63–76, 10.1016/j.blre.2016.08.005. [DOI] [PubMed] [Google Scholar]
- 9. Palomo L., Malinverni R., Cabezón M., et al., “DNA Methylation Profile in Chronic Myelomonocytic Leukemia Associates With Distinct Clinical, Biological and Genetic Features,” Epigenetics 13, no. 1 (2018): 8–18, 10.1080/15592294.2017.1405199. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Olarte Carrillo I., Ramos Peñafiel C., Miranda Peralta E., et al., “Clinical Significance of the ABCB1 and ABCG2 Gene Expression Levels in Acute Lymphoblastic Leukemia,” Hematology 22, no. 5 (2016): 286–291, 10.1080/10245332.2016.1265780. [DOI] [PubMed] [Google Scholar]
- 11. George T. I., “Malignant or Benign Leukocytosis,” Hematology 2012, no. 1 (2012): 475–484, 10.1182/asheducation.v2012.1.475.3798515. [DOI] [PubMed] [Google Scholar]
- 12. Chabot‐Richards D. S. and George T. I., “Leukocytosis,” International Journal of Laboratory Hematology 36, no. 3 (2014): 279–288, 10.1111/ijlh.12212. [DOI] [PubMed] [Google Scholar]
- 13. Renieblas G. P., Nogués A. T., González A. M., Gómez‐Leon N., and Del Castillo E. G., “Structural Similarity Index Family for Image Quality Assessment in Radiological Images,” Journal of Medical Imaging 4, no. 3 (2017): e035501, 10.1117/1.JMI.4.3.035501. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Geneviève F., Godon A., Marteau‐Tessier A., and Zandecki M., “Anomalies et Erreurs de Détermination de L'hémogramme Avec les Automates D'hématologie Cellulaire Partie 2. Numération et Formule Leucocytaires,” Annales de Biologie Clinique 70, no. 2 (2012): 141–154, 10.1684/abc.2012.0665. [DOI] [PubMed] [Google Scholar]
- 15. Osman M., Akkus Z., Jevremovic D., et al., “Classification of Monocytes, Promonocytes and Monoblasts Using Deep Neural Network Models: An Area of Unmet Need in Diagnostic Hematopathology,” Journal of Clinical Medicine 10, no. 11 (2021): 2264, 10.3390/jcm10112264. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Sidhom J.‐W., Siddarthan I. J., Lai B.‐S., et al., “Deep Learning for Diagnosis of Acute Promyelocytic Leukemia via Recognition of Genomically Imprinted Morphologic Features,” npj Precision Oncology 5, no. 1 (2021): 38, 10.1038/s41698-021-00179-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Huang F., Guang P., Li F., Liu X., Zhang W., and Huang W., “AML, ALL, and CML Classification and Diagnosis Based on Bone Marrow Cell Morphology Combined With Convolutional Neural Network: A STARD Compliant Diagnosis Research,” Medicine 99, no. 45 (2020): e23154, 10.1097/MD.0000000000023154. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Lapsina S., Riond B., Hofmann‐Lehmann R., and Stirn M., “Comparison of Sysmex XN‐V Body Fluid Mode and Deep‐Learning‐Based Quantification With Manual Techniques for Total Nucleated Cell Count and Differential Count for Equine Bronchoalveolar Lavage Samples,” BMC Veterinary Research 20, no. 1 (2024): 48, 10.1186/s12917-024-03884-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Alcazer V., Le Meur G., Roccon M., et al., “Evaluation of a Machine‐Learning Model Based on Laboratory Parameters for the Prediction of Acute Leukaemia Subtypes: A Multicentre Model Development and Validation Study in France,” Lancet Digital Health 6, no. 5 (2024): 323–333, 10.1016/s2589-7500(24)00044-x. [DOI] [PubMed] [Google Scholar]
- 20. Liao H., Xu Y., Meng Q., et al., “A Convolutional Neural Network‐Based, Quantitative Complete Blood Count Scattergram‐Mapping Framework Promptly Screens Acute Promyelocytic Leukemia With High Sensitivity,” Cancer 129, no. 19 (2023): 2986–2998, 10.1002/cncr.34890. [DOI] [PubMed] [Google Scholar]
- 21. Li Z., Fan Q., Bilgic B., et al., “Diffusion MRI Data Analysis Assisted by Deep Learning Synthesized Anatomical Images (DeepAnat),” Medical Image Analysis 86 (2023): 102744, 10.1016/j.media.2023.102744. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Talo M., Yildirim O., Baloglu U. B., Aydin G., and Acharya U. R., “Convolutional Neural Networks for Multi‐Class Brain Disease Detection Using MRI Images,” Computerized Medical Imaging and Graphics 78 (2019): 101673, 10.1016/j.compmedimag.2019.101673. [DOI] [PubMed] [Google Scholar]
- 23. Khan R. A., Fu M., Burbridge B., Luo Y., and Wu F.‐X., “A Multi‐Modal Deep Neural Network for Multi‐Class Liver Cancer Diagnosis,” Neural Networks 165 (2023): 553–561, 10.1016/j.neunet.2023.06.013. [DOI] [PubMed] [Google Scholar]
- 24. Yang Y., Zhang L., Du M., et al., “A Comparative Analysis of Eleven Neural Networks Architectures for Small Datasets of Lung Images of COVID‐19 Patients Toward Improved Clinical Decisions,” Computers in Biology and Medicine 139 (2021): 104887, 10.1016/j.compbiomed.2021.104887. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Stiller S., Dueñas J. F., Hempel S., Rillig M. C., and Ryo M., “Deep Learning Image Analysis for Filamentous Fungi Taxonomic Classification: Dealing With Small Datasets With Class Imbalance and Hierarchical Grouping,” Biological Methods and Protocols 9, no. 1 (2024): bpae063, 10.1093/biomethods/bpae063. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Hirahara D., Takaya E., Takahara T., and Ueda T., “Effects of Data Count and Image Scaling on Deep Learning Training,” PeerJ Computer Science 6 (2020): e312, 10.7717/peerj-cs.312. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Fu B., Zhang M., He J., Cao Y., Guo Y., and Wang R., “StoHisNet: A Hybrid Multi‐Classification Model With CNN and Transformer for Gastric Pathology Images,” Computer Methods and Programs in Biomedicine 221 (2022): 106924, 10.1016/j.cmpb.2022.106924. [DOI] [PubMed] [Google Scholar]
- 28. Naresh K. N. and Medeiros L. J., “Introduction to the Fifth Edition of the World Health Organization Classification of Tumors of Hematopoietic and Lymphoid Tissues,” Modern Pathology 36, no. 12 (2023): 100330, 10.1016/j.modpat.2023.100330. [DOI] [PubMed] [Google Scholar]
- 29. Inaba T., Ishizuka K., Yuasa S., et al., “Abnormal Neutrophil Scattergram Obtained Using Pentra MS CRP in the Patients With Myelodysplastic Syndrome Showing Dysgranulopoiesis,” International Journal of Laboratory Hematology 38, no. 1 (2016): 27–33, 10.1111/ijlh.12423. [DOI] [PubMed] [Google Scholar]
- 30. Kiya G. T., Aklilu Getachew M., Biya S. A., et al., “Validation of the International Consensus Group Criteria for Slide Review Following Automated Complete Blood Count at Jimma Medical Center, Ethiopia,” Journal of Blood Medicine 14 (2023): 213–220, 10.2147/JBM.S402940. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Cai Q., Ye B., Zheng W., et al., “Development of a Screening Model for APL Using Cell Population Data and Deep Learning‐Extracted WBC Scattergram Features,” BMC Cancer 25, no. 1 (2025): 1725, 10.1186/s12885-025-15034-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Zhang Y., Hong D., McClement D., Oladosu O., Pridham G., and Slaney G., “Grad‐CAM Helps Interpret the Deep Learning Models Trained to Classify Multiple Sclerosis Types Using Clinical Brain Magnetic Resonance Imaging,” Journal of Neuroscience Methods 353 (2021): 109098, 10.1016/j.jneumeth.2021.109098. [DOI] [PubMed] [Google Scholar]
- 33. Alharbi A. H., Aravinda C. V., Lin M., Ashwini B., Jabarulla M. Y., and Shah M. A., “Detection of Peripheral Malarial Parasites in Blood Smears Using Deep Learning Models,” Computational Intelligence and Neuroscience 2022, no. 1 (2022): 3922763, 10.1155/2022/3922763. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Rodellar J., Alférez S., Acevedo A., Molina A., and Merino A., “Image Processing and Machine Learning in the Morphological Analysis of Blood Cells,” International Journal of Laboratory Hematology 40, no. S1 (2018): 46–53, 10.1111/ijlh.12818. [DOI] [PubMed] [Google Scholar]
- 35. Del Casale A., Sarli G., Bargagna P., et al., “Machine Learning and Pharmacogenomics at the Time of Precision Psychiatry,” Current Neuropharmacology 21, no. 12 (2023): 2395–2408, 10.2174/1570159X21666230808170123. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Zhu S., Li Y., Zhang F., et al., “Raman Spectromics Method for Fast and Label‐Free Genotype Screening,” Biomedical Optics Express 14, no. 6 (2023): 3072–3085, 10.1364/BOE.493524. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37. Cohen J. P., Cao T., Viviano J. D., et al., “Problems in the Deployment of Machine‐Learned Models in Health Care,” CMAJ 193, no. 35 (2021): E1391–E1394, 10.1503/cmaj.202066. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Xu N., Wang J., Dai G., et al., “EfficientNet‐Based System for Detecting Egfr‐Mutant Status and Predicting Prognosis of Tyrosine Kinase Inhibitors in Patients With NSCLC,” Journal of Imaging Informatics in Medicine 37, no. 3 (2024): 1086–1099, 10.1007/s10278-024-01022-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39. Li A., Li M., Fei R., Mallik S., Hu B., and Yu Y., “EfficientNet‐resDDSC: A Hybrid Deep Learning Model Integrating Residual Blocks and Dilated Convolutions for Inferring Gene Causality in Single‐Cell Data,” Interdisciplinary Sciences: Computational LIfe Sciences 17, no. 1 (2024): 166–184, 10.1007/s12539-024-00667-2. [DOI] [PubMed] [Google Scholar]
- 40. Xu W., Nie L., Chen B., and Ding W., “Dual‐Stream EfficientNet With Adversarial Sample Augmentation for COVID‐19 Computer Aided Diagnosis,” Computers in Biology and Medicine 165 (2023): 107451, 10.1016/j.compbiomed.2023.107451. [DOI] [PubMed] [Google Scholar]
- 41. Huang C., Wang W., Zhang X., Wang S.‐H., and Zhang Y.‐D., “Tuberculosis Diagnosis Using Deep Transferred Efficientnet,” IEEE/ACM Transactions on Computational Biology and Bioinformatics 20, no. 5 (2023): 2639–2646, 10.1109/tcbb.2022.3199572. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Figure S1: STARD‐style flow of WBC‐Diff scattergram images. The development cohort (n = 4297; 2016–2024) was split at the image level into training (n = 2568), validation (n = 863), and test (n = 866) sets (stratified; 60/20/20; random_state = 42). Offline augmentation was applied to the training set only, expanding it to 7956 images. An independent hold‐out cohort (n = 500; 2025), disjoint from the development cohort at both the image and patient level, was evaluated without any retraining or fine‐tuning. Validation, test, and independent hold‐out sets comprised original (non‐augmented) images only. A content‐hash (MD5) plus acquisition‐metadata audit confirmed no image‐ or patient‐level overlap across splits (File S1).
Table S1: The account of initial, expand quantity of augmentation, and total number in each class.
Table S2: The count of each class for dataset split model training, validation, and evaluation.
Table S3: Overfit detection: Comparison of validation set and test set performance for EfficientNet‐B0 multi‐class classification.
Table S4: Post hoc pairwise comparisons using Tukey's HSD test on PCA components derived from EfficientNet‐B0 Grad‐CAM Activation Vectors.
Table S5: Post hoc pairwise comparisons using Tukey's HSD test on PCA components derived from EfficientNet‐B0 Grad‐CAM activation vectors.
Table S6: Patient‐level data audit for the four rare hematological classes.
File S1: Data provenance and split audit report.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author reasonable request. The data are not publicly available because of privacy or ethical restrictions.