Automated Hybrid Feature Extraction for White Blood Cells Classification Using Radiomics and Graph Neural Networks (GNN)

Abstract

White Blood Cell (WBC) classification plays a crucial role in diagnosing various blood-related diseases. However, the manual classification of WBCs from microscopic images is time-consuming, prone to human error, and lacks consistency. Existing automated methods primarily rely on Convolutional Neural Networks (CNNs), but they struggle with accurately classifying cells that have complex morphological structures or are adhered together. To address these limitations, we propose an Automated Hybrid Feature Extraction System combining Radiomics and Graph Neural Networks (GNN) for accurate WBC classification. The system integrates detailed radiomic features (shape, color, texture) and spatial relationships between features modeled by GNN, resulting in improved classification accuracy. Our model is capable of handling complex cellular structures, outperforming traditional CNN-based approaches. In our experiments, the hybrid model achieves 96% accuracy on the validation set, significantly enhancing the precision and reliability of WBC classification for medical diagnosis. This research offers a novel approach to improving the diagnostic capabilities of automated WBC classification systems.