Advanced Deep Learning for White Blood Cell Classification: A Hybrid and Interpretable Approach
International Research Journal of Modernization in Engineering Technology and Science · 2024
White blood cell (WBC) classification is essential for diagnosing hematological disorders such as leukemia and infections.Despite advancements in deep learning, challenges like class imbalance, small dataset sizes, and lack of interpretability persist.This paper introduces a novel ensemble framework that combines EfficientNet, Vision Transformers (ViT), DenseNet, ResNet, and a custom lightweight convolutional neural network (CNN) to enhance classification accuracy and robustness.To address class imbalance, Generative Adversarial Networks (GANs) generate synthetic samples for underrepresented classes, while Focal Loss emphasizes hard-to-classify instances.The ensemble integrates predictions using weighted averaging, leveraging the strengths of each architecture.Interpretability is achieved through Grad-CAM and SHAP, providing visual and quantitative insights into model predictions, and fostering clinical trust.The framework achieves a test accuracy of 85.44% on a benchmark WBC dataset, outperforming individual models such as ResNet (82.9%) and DenseNet (83.4%).This research demonstrates a strong and interpretable solution for WBC classification, overcoming significant weaknesses in current approaches.The proposed framework holds significant potential for improving diagnostic workflows in clinical settings.