Clinical Application of Using Diffusion‐Based Wasserstein Generative Adversarial Network for Morphologic Analysis of Blood Cells

ABSTRACT

Background

Morphologic analysis of peripheral blood smears is essential for diagnosing hematologic diseases and patient management. Although manual microscopy is the traditional gold standard, it is time‐consuming and subjective. Digital morphology analyzers have improved automation and accuracy; however, challenges remain, particularly in classifying certain cell types. Recently, the diffusion‐based Wasserstein generative adversarial network with gradient penalty (DWGAN‐GP) showed potential by enhancing image quality and addressing data imbalance. We aim to investigate the accuracy of blood cell classification using the DWGAN‐GP model.

Methods

In this study, the DWGAN‐GP model in conjunction with the EfficientNetB3 classification model was evaluated using 78,494 peripheral blood cell images. Samples were collected from patients with normal and abnormal hematologic conditions. Data were balanced by augmenting underrepresented classes with synthetic images, resulting in equal representation across 13 cell classes. Performance was compared with PBIA (ANI Co., Suwon, Korea), a commercial digital morphology analyzer.

Results

DWGAN‐GP augmentation significantly improved classification accuracy of the EfficientNetB3 model to 97.74% with an F1‐score of 91.13%. This result surpassed both the unbalanced dataset (accuracy 95.68%, F1‐score 82.12%) and PBIA system (accuracy 95%). Notably, improvements were significant in minority classes such as blasts and myelocytes, which are critical in diagnosing leukemia.

Conclusion

Incorporating synthetic data using DWGAN‐GP significantly enhanced model performance and addressed class imbalance. This method shows promise for more accurate and consistent blood cell classification, offering potential improvements in clinical diagnostics for hematologic disorders.

This study evaluated the accuracy of blood cell classification using 78,494 images across 13 cell classes, demonstrating a significant improvement with the DWGAN‐GP augmented dataset. The use of synthetic images increased the classification accuracy to 97.74%, outperforming both unbalanced data and commercial systems. These results highlight that incorporating DWGAN‐GP greatly enhances the precision and reliability of blood cell classification in hematologic diagnostics.

1. Introduction

Morphologic analysis of the peripheral blood smear is an essential component in diagnosing hematologic diseases and in patient management [1]. Traditionally, the morphologic analysis of the blood smear has been performed using manual microscopic examination for several decades. Although this method is considered the gold standard, it is a time‐consuming and labor‐intensive procedure that requires skilled technicians and is subject to considerable interobserver variability due to substantial reviewer subjectivity [2].

Many advanced digital morphology analyzers, such as CellaVision (CellaVision AB, Lund, Sweden), DI‐60 (Sysmex, Kobe, Japan), MC‐80 (Mindray Medical International Ltd., Shenzhen, China), and UIMD PBIA (ANI CO., Suwon, Korea), have been released to meet the increasing demand for automated morphologic analysis of peripheral blood smears [3, 4, 5, 6]. Depending on the digital morphology analyzer and conditions of the evaluated samples and images, the overall classification accuracy ranges from 82% to 99% [5, 6, 7]. The accuracy is particularly high for differentiating white blood cells such as neutrophils, lymphocytes, and monocytes. However, misclassification frequently occurs in immature granulocytes, blasts, and abnormal lymphocytes, with reported accuracies between 81% and 94% [6]. These analyzers locate cells in the blood smear and capture individual blood cell images at high magnification using a camera. Digital images of individual cells are then used as input for computer‐aided classification using a neural network based on a large database of cells utilizing geometric, color, and texture features [8].

Recent advancements in deep learning models for blood cell image analysis have been reported. Notably, the new diffusion‐based Wasserstein generative adversarial network with gradient penalty (DWGAN‐GP) is a generative model that offers high‐quality images for rare blood cell classes and addresses data imbalance [9]. This approach reportedly speeds up generation and improves output fidelity, achieving a 95% average accuracy, especially for complex blood cell images. However, the performance demonstrated in the previous study was limited because only 4500 images collected from a single institution were used.

To overcome the data limitations that often lead to misclassification, especially for rare cell types, various data augmentation techniques have been developed. While traditional methods like rotation, scaling, and flipping can increase dataset size, they offer limited novelty and often fail to capture the complex morphological diversity of biological samples. More advanced approaches have turned to generative models to create realistic, high‐quality synthetic data. Generative Adversarial Networks (GANs), for instance, have been used to augment thermographic images for defect detection by generating new samples that enlarge the diversity of the original dataset [10]. However, GANs can be difficult to train and may suffer from issues like model collapse, where the variety of generated samples is limited [9]. More recently, Denoising Diffusion Probabilistic Models (DDPMs) have shown exceptional performance in generating high‐quality data, even from small sample sizes, by progressively removing noise from a random signal [11]. Yet, this high quality often comes at the cost of slow sampling and high computational demands [9]. Our approach utilizes the DWGAN‐GP model [9], which was specifically designed to balance these trade‐offs by combining the stable training of a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN‐GP) with the structured noise input from a DDPM's forward process, enabling the rapid generation of diverse, high‐fidelity images suited for complex clinical datasets.

In this study, the accuracy of blood cell classification was validated using the DWGAN‐GP model for generating synthetic blood cell images to balance the dataset of 78,494 images from patients with normal or abnormal clinical conditions. In addition, the performance of the DWGAN‐GP‐assisted model and the commercial digital morphology analyzer PBIA was compared.

2. Methods

2.1. Study Samples

This study was performed using routine complete blood count (CBC) with differential samples collected from August to September 2023. The dataset comprised 356 samples, including 190 with abnormal findings (53.4%) and 166 from normal subjects (46.6%). From these samples, the differential percentage (%) and peripheral blood cell images were obtained. Among the abnormal samples, 36 (18.9%) were from acute leukemia patients, with 21 (11.1%) with acute myeloid leukemia, 14 (7.4%) with acute lymphoblastic leukemia, and one (0.5%) with acute undifferentiated leukemia. In addition, 36 samples (18.9%) were from lymphoma patients, consisting of seven (3.7%) with chronic lymphocytic leukemia, 25 (13.2%) with other B cell lymphomas, three (1.6%) with T cell lymphoma, and one (0.5%) with Hodgkin lymphoma. There were also nine samples (4.7%) from multiple myeloma patients, 12 (6.3%) from myelodysplastic neoplasm patients, eight (4.2%) from chronic myeloid leukemia patients, eight (4.2%) from primary myelofibrosis patients, and two (1.1%) from essential thrombocythemia patients. The dataset also included 28 samples (14.7%) from solid cancer patients and 51 (26.8%) from patients with other underlying diseases. The remaining 166 were collected from healthy individuals undergoing routine health check‐ups. The study protocol was reviewed and approved by the Institutional Review Board of the Samsung Medical Center (IRB no. 2024‐07‐119). The appropriateness of the purpose for utilizing the research data and the adequacy of the anonymization method were reviewed and approved by the Research Data Review Committee of the Samsung Medical Center (DRB no. D2024‐0091). This study adhered to the tenets of the Declaration of Helsinki. Given the retrospective nature of the study and the use of anonymized data, the requirement for informed consent was waived. No images or video or details that could identify the subjects were used in the study. The data used in this study was accessed on December 20, 2024, after it was collected.

2.2. Peripheral Blood Cell Analysis

Peripheral blood samples were collected into K₂ EDTA tubes (Becton, Dickinson and Company, NJ, USA) and analyzed within 4 h after blood collection using the Sysmex XN‐9000 hematology analyzer (Sysmex, Kobe, Japan). Blood smears were prepared using the Sysmex SP‐100 slide maker and stainer (Sysmex) in an automated hematology slide preparation unit. Blood cell images were then acquired using the PBIA digital image analyzer. Initially, 17 types of blood cells were automatically classified using the UPBA‐12A classification software (ANI CO.). The types included segmented neutrophil (NES), band neutrophil (NEB), metamyelocyte (ME), myelocyte (MY), promyelocyte (PR), lymphocyte (LY), lymphocyte reactive (LR), lymphocyte abnormal, monocyte (MO), eosinophil (EO), basophil (BA), blast (BL), plasma cell (PC), nucleated RBC (NR), giant platelet (GP), platelet aggregation, and artifact (AR). Following this automated classification, experts manually reviewed and reclassified the cells.

2.3. Data Exploration and Preprocessing

A classification evaluation was conducted on 78,494 blood cell images (240 × 240 pixels, 24‐bit BMP) across 15 blood cell classes. This classification integrates previously reported cell classes [9] with PBIA cell classes, incorporating several modifications: the large granular lymphocyte class was removed, artifact and smudge cell classes were consolidated into AR, and a new PC class was added. The following 15 classes were evaluated: AR, BA, BL, EO, GP, lymphocyte variant form (LR), LY, ME, MO, MY, NEB, NES, NR, PC, and PR. The distribution of the original dataset is detailed in Table S1. Due to the constraints of data availability, the focus in the study was on 13 classes, omitting PC and PR from the analysis.

The dataset was evenly split into training and testing sets (50:50 ratio) to ensure robust performance metrics and model generalization. This balanced division provides sufficient data for both training and comprehensive testing, enhancing the model's stability and reliability. To address class imbalance, a generative model (DWGAN‐GP) was used to augment training data for underrepresented classes to 3600 images per class. We chose the DWGAN‐GP model specifically to overcome the common limitations of other generative models. In practice, standard GANs often suffer from training instability and mode collapse, which makes it difficult to generate a diverse set of high‐quality images for rare classes. Our model first addresses this by incorporating the WGAN‐GP framework, which uses a Wasserstein loss with a gradient penalty to ensure a stable training process and prevent the generator from producing repetitive, low‐variety samples. Secondly, while a standalone diffusion model can generate excellent images, their clinical utility is limited by an extremely slow and computationally expensive sampling process. Our hybrid DWGAN‐GP model solves this efficiency problem by replacing the slow, multistep reverse diffusion process with a single, efficient pass through a GAN generator. By using structured noisy vectors from the DDPM's forward process as input instead of simple random noise, we accelerate model convergence and gain finer control over the output, which is critical for faithfully reproducing the complex morphologies of clinical blood cells. This approach ensured equal representation of all classes during training, mitigating potential biases, and improving the ability of the model to generalize across diverse cell types. Classes with abundant data, such as AR, LY, and NES (9943, 8474, and 14,604 images, respectively), were not augmented because they already contained sufficient samples for effective training. Initially, the training set comprised 39,231 images, with 39,238 images reserved for testing. The generative model contributed an additional 29,790 synthetic images, resulting in a balanced training set of 69,021 images. This process in a pipeline diagram is shown in Figure 1, depicting the workflow from initial dataset preparation through generative model training, synthetic data generation, class balancing, and final classification model training. The training process incorporated standard data augmentation techniques, including rotation, scaling, and flipping, to further enhance model generalization. The final performance of the model was evaluated using the reserved test set to assess accuracy in classifying various blood cell types.

FIGURE 1.

Data and model pipeline of a deep learning classification model enhanced using a generative model for data augmentation: (X) Input data, (g) Generative model (diffusion‐based Wasserstein generative adversarial network with gradient penalty, DWGAN‐GP), (e) Classification model (EfficientNetB3).

2.3.1. DWGAN‐GP Data Augmentation and Training

The generative pipeline of the DWGAN‐GP model combines (i) the forward diffusion process from a DDPM to construct structured noise vectors and (ii) a WGAN‐GP generator–critic pair trained with a gradient‐penalty objective for stable image synthesis.

For each minority class, we applied the DDPM forward process to its real images to generate structured noisy image vectors that served as inputs to the GAN generator (instead of standard Gaussian noise). We used a linear schedule with 1000 diffusion steps (coefficient 0.01), which we previously found provides a good trade‐off between model expressiveness and compute.

The WGAN‐GP critic and generator each comprised four 2D convolutional blocks with Leaky‐ReLU activations and batch normalization to improve speed, stability, and performance. We trained with a small batch size, a learning rate of 0.0002, and ran 8000 epochs. The gradient penalty was implemented via a Keras RandomWeightedAverage layer to compute the penalty on interpolated samples, which helps enforce the Lipschitz constraint and prevents mode collapse. To avoid biasing toward majority patterns, image generation was performed class‐wise for each minority class. Generated samples were visually screened to ensure variety and realism before inclusion in training. EfficientNetB3 was trained on the balanced (combined real and synthetic) training set with standard data augmentations (rotation, scaling, flipping). Model selection and evaluation used the held‐out test set, and summary metrics reported in Section 3 are averaged across folds in our experiments.

2.4. Statistics

To assess the performance of the model for blood cell classification, overall accuracy and macro‐average metrics were utilized. The macro‐average metrics offer a balanced approach to evaluating multi‐class classification performance by computing the precision, recall, and F1‐score (overall prediction accuracy) for each class individually and then averaging these scores across all classes [12].

3. Results

3.1. Enhanced Blood Cell Classification Performance Using DWGAN‐GP Data Augmentation

The integration of synthetic data generated when using the DWGAN‐GP model significantly enhanced the balance of the training dataset, leading to an improved performance of blood cell classification by EfficientNetB3 [13] (Table 1). This improvement was evident when comparing the average metric performance across 10 folds for both the original unbalanced and the augmented balanced datasets. Analysis of the original unbalanced dataset revealed a substantial decrease in performance metrics from training to test sets. The test accuracy was 0.9568 with an F1‐score of 0.8212, indicating suboptimal generalization capabilities. In contrast, the augmented balanced dataset demonstrated superior performance across both training and test sets. The test accuracy increased to 0.9774 and the F1‐score improved to 0.9113. These results indicated enhanced generalization and model robustness. The balanced training data facilitated more effective learning across all classes, mitigating bias and improving classification accuracy for various blood cell types.

TABLE 1.

Comparison of blood cell classification performance with and without DWGAN‐GP data augmentation.

Measure	Average metric performance on 10‐folds
	Training performance	Testing performance
Original unbalanced dataset
Accuracy	0.9861	0.9568
Precision	0.9854	0.8714
Recall	0.9841	0.8187
F1‐score	0.9847	0.8212
Augmented balanced dataset with DWGAN‐GP
Accuracy	0.9883	0.9774
Precision	0.9785	0.9350
Recall	0.9882	0.8980
F1‐score	0.9824	0.9113

Note: The EfficientNetB3 model was used for blood cell classification. Training performance: model evaluation on training dataset. Testing performance: model evaluation on the unseen testing dataset.

Abbreviation: DWGAN‐GP, diffusion‐based Wasserstein generative adversarial network with gradient penalty.

3.2. Comparison of Classification Accuracy Between the DWGAN‐GP‐Assisted Model and the Commercial Digital Morphology Analyzer PBIA System

Three approaches to blood cell classification were evaluated using macro‐average and general classification accuracy metrics. The PBIA system achieved a macro‐average classification accuracy of 0.86 and a general classification accuracy of 0.95. Utilizing the EfficientNetB3 model with the original dataset yielded improved results compared with the PBIA system, attaining a macro‐average classification accuracy of 0.82 and a general classification accuracy of 0.96. The integration of DWGAN‐GP augmented data with the EfficientNetB3 model demonstrated the most substantial performance enhancement, achieving a macro‐average classification accuracy of 0.91 and a general classification accuracy of 0.98. This augmented approach notably outperformed both the PBIA system and the original EfficientNetB3 model, particularly in terms of macro‐average classification accuracy, indicating improved performance across all cell classes.

3.3. Class‐Wise Performance of the DWGAN‐GP‐Augmented EfficientNetB3 Model and the PBIA System

The DWGAN‐GP‐augmented EfficientNetB3 model demonstrated superior performance across most cell classes compared with the PBIA system, particularly in classes with limited original training samples (Figure 2 and Table S2). Significant improvements were observed in classes where PBIA exhibited limitations, notably in BL and MY classes. However, marginal accuracy decreases in certain classes, such as GP, indicated areas for potential model enhancement.

FIGURE 2.

Confusion table for white blood cells (total N = 78,494). (A) Class‐wise accuracy distribution and (B) number of predicted cells per class. Diffusion‐based Wasserstein generative adversarial network with gradient penalty (DWGAN‐GP)‐augmented EfficientNetB3 classification (x‐axis, “Predicted Label”) were compared with expert re‐classification (y‐axis, “True Label”).

3.4. Enhanced Classification Through Augmented Visualizations

Our study demonstrates that augmenting training data with synthetic images significantly improves classification performance. We provided a qualitative image comparison (Figure 3), illustrating original and DWGAN‐GP generated synthetic images for key minority classes such as basophil and lymphocyte variant forms. These synthetic images exhibit high fidelity, successfully replicating critical morphological features. Moreover, a t‐distributed stochastic neighbor embedding (t‐SNE) visualization (Figure 4) reveals that synthetic data points closely align with real data within the EfficientNetB3 classifier's feature space. This alignment indicates that our synthetic images not only appear realistic but also serve as effective training examples, thus explaining the observed enhancement in classification performance.

FIGURE 3.

Qualitative comparison of original and synthetic blood cell images for selected minority classes. The top row presents examples of original, real cell images for basophil (BA), giant platelet (GP), lymphocyte variant form (LR), metamyelocyte (ME), myelocyte (MY), and nucleated RBC (NR). The bottom row displays samples of corresponding high‐fidelity synthetic images generated by the DWGAN‐GP model. The generated images successfully replicate key morphological features, including nuclear shape, cytoplasm color, and granularity, demonstrating the model's effectiveness for data augmentation.

FIGURE 4.

T‐distributed stochastic neighbor embedding (t‐SNE) visualization of learned features for real and synthetic data. The plot shows the 2D t‐SNE embedding of feature vectors extracted from the penultimate layer of the EfficientNetB3 classifier. Each point corresponds to an image, with circles representing real images and triangles representing synthetic images generated by the DWGAN‐GP model. The close clustering and intermingling of real and synthetic points within each class (indicated by color) demonstrate that the classification model learns similar feature representations for both, validating the high quality and effectiveness of the augmented data.

4. Discussion

In this study, the clinical applicability of the DWGAN‐GP‐augmented EfficientNetB3 model for blood cell classification was evaluated in a clinical laboratory using over 78,000 images obtained from patients with various hematological malignancies. This study is the first in which a deep learning classification model for data augmentation was demonstrated to be useful for analyzing various blood cells, including abnormal cells.

This work provides several important findings. First, the implementation of synthetic data generated when using the DWGAN‐GP model significantly enhanced the performance metrics of the EfficientNetB3 model. Specifically, the augmented model achieved a test accuracy of 97.74% and an F1‐score of 91.13%, markedly surpassing the performance of the original unbalanced dataset (95.68% accuracy and 82.12% F1‐score). The augmented balanced dataset facilitated more effective learning across all blood cell classes, thereby reducing bias and enhancing the generalization capabilities of the model. When compared with the PBIA system, the DWGAN‐GP‐augmented EfficientNetB3 model demonstrated superior performance. Although the PBIA system achieved a general classification accuracy of 95% and a macro‐average classification accuracy of 86%, our model attained 98% and 91%, respectively. These results indicated better overall performance and more consistent accuracy across diverse cell classes, a crucial factor in medical diagnostics.

Second, the augmented model showed an improved ability to generalize, especially in minority classes, providing a more accurate and consistent tool for blood cell classification. Notably, in the classification of the BL class, which is clinically significant and challenging to learn due to its varied morphological characteristics [1, 2], our model achieved an accuracy of 92%, significantly higher than 66% with the PBIA system. Similarly, the augmented model achieved 92% accuracy in the MY class, compared with 70% with the PBIA system. BL and MY are blood cells that can be observed in patients with various hematological malignancies, including leukemia [14, 15]. A morphological blood cell assessment represents the first step in the diagnostic pathway for the primary diagnosis of leukemia. However, these cells can exhibit different morphology in peripheral blood compared with bone marrow, and are heterogeneous in shape and size, limiting morphological evaluation [16, 17]. The capacity to overcome class imbalance and improve overall classification accuracy, particularly in underrepresented classes, highlights the potential of the model for practical applications in medical diagnostics and as an advanced solution for blood cell classification in clinical settings.

The PBIA system is an artificial intelligence algorithm‐based cell image classification model and requires a training process using images of various blood cells to improve classification accuracy [6]. In this study, the accuracy of blood cell classification may have been affected because the PBIA system used for evaluation included results both from the default model before training and the upgraded model after training. Although the accuracy of the PBIA system in cell classification can improve with extended learning, this process is time‐consuming. The approach proposed in this study could be particularly beneficial for rapid clinical test setup, especially when acquiring a large image dataset initially is challenging.

In conclusion, these findings of this study showed the potential of integrating deep learning models with sophisticated data augmentation techniques in medical image classification. This approach improves overall accuracy and addresses the critical issue of class imbalance, providing the basis for creating more reliable and consistent diagnostic tools in clinical hematology.

Author Contributions

Conceptualization and methodology: Hyun‐Young Kim and Mi‐Ae Jang. Formal analysis: Emmanuel Edward Ngasa, Hyun‐Young Kim, and Jiyoung Woo. Sample collection and investigation: Hee‐Jin Kim, Boram Kim, Gyujin Lim, Chang‐Hun Park, and Hyeon Jeong Kwon. Supervision: Mi‐Ae Jang and Jiyoung Woo. Writing – original draft: Emmanuel Edward Ngasa, Hyun‐Young Kim, and Jiyoung Woo. Writing – review and editing: all authors.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Tables S1–S2: jcla70118‐sup‐0001‐TablesS1‐S2.docx.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.