Combining DC-GAN with ResNet for blood cell image classification.

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• White blood cell (WBC) classification is crucial for diagnosing various diseases.
• Traditional methods face limitations in accuracy due to segmentation dependencies and insufficient/unbalanced data.
• Deep learning models can struggle with data scarcity in medical image analysis.

Purpose of the Study:

• To develop a robust framework for accurate white blood cell image classification.
• To address challenges of insufficient data and improve deep learning model performance in medical diagnosis.
• To enhance the discriminative power of learned features for better classification.

Main Methods:

• Proposed a novel framework combining Deep Convolutional Generative Adversarial Networks (DC-GAN) with Residual Neural Networks (ResNet).
• Utilized DC-GAN for generating synthetic WBC images to supplement training data.
• Introduced a modified loss function to increase inter-class variation and decrease intra-class differences, improving feature discriminability.

Main Results:

• The proposed model achieved a high classification accuracy of 91.7% for WBC images.
• The integration of DC-GAN generated samples improved the overall classification performance.
• The modified loss function enhanced the model's ability to distinguish between different WBC types.

Conclusions:

• The developed DC-GAN and ResNet framework offers a significant improvement in WBC image classification accuracy.
• Data augmentation using generative adversarial networks is effective in overcoming data limitations in medical imaging.
• The novel loss function contributes to more discriminative feature learning for accurate medical diagnosis.