Multi-Scale Hematological Image Analysis for WBC Classification via GNNs

Abstract

Traditional deep learning models, like the Convolutional Neural Networks (CNNs) fails in domain shift due to inconsistencies arising due to variances in staining and imaging techniques. This study attempts to use Graph Neural Networks (Graph Neural Networks) for the classification of WBCs given their ability to detect structural and relational patterns from microscopic images. Now, to avoid redundancy and allow generalization, out of 47 datasets six datasets plus the combined set were chosen. The technique includes image preprocessing, normalization, resizing to 224×224 pixels, and data enhancement to promote more robustness in the model. The GNN model was trained and tested utilizing accuracy metrics, precision metrics, recall, F1-score, confusion matrices, and ROC curves. Comparative analysis indicated that by training the model on a heterogeneous dataset, enhances model generalization, and identifying spatial dependencies in the WBC images was done better than with CNNs. These results demonstrate the potential of Graph Neural Networks in AI-based hematological diagnostics to further strengthen medical image analysis.