Clinical Application of Using Diffusion-Based Wasserstein Generative Adversarial Network for Morphologic Analysis of...

Area of Science:

• Hematology
• Artificial Intelligence
• Medical Imaging

Background:

• Manual microscopy for blood smear analysis is time-consuming and subjective.
• Digital morphology analyzers offer automation but face challenges in classifying certain cell types.
• Diffusion-based Wasserstein generative adversarial network with gradient penalty (DWGAN-GP) shows potential for image enhancement and data balancing.

Purpose of the Study:

• To investigate the accuracy of blood cell classification using the DWGAN-GP model.
• To evaluate the effectiveness of DWGAN-GP in enhancing image quality and addressing data imbalance in peripheral blood cell analysis.

Main Methods:

• The DWGAN-GP model was combined with the EfficientNetB3 classification model.
• 78,494 peripheral blood cell images from patients with normal and abnormal hematologic conditions were used.
• Data augmentation with synthetic images balanced underrepresented classes, creating equal representation across 13 cell types.

Main Results:

• DWGAN-GP augmentation improved EfficientNetB3 classification accuracy to 97.74% (F1-score 91.13%), surpassing the unbalanced dataset (95.68% accuracy, 82.12% F1-score) and a commercial analyzer (95% accuracy).
• Significant improvements were observed in classifying minority cell types, including blasts and myelocytes, crucial for leukemia diagnosis.
• The synthetic data approach effectively addressed class imbalance, leading to more robust classification.

Conclusions:

• Integrating synthetic data via DWGAN-GP significantly boosts model performance and resolves class imbalance issues.
• This AI-driven method offers a promising approach for more accurate and consistent blood cell classification.
• The findings suggest potential improvements in clinical diagnostics for various hematologic disorders.