XHSA-DCNet: An Explainable Hybrid Swin Transformer and Attention-Guided Dense Convolution Network for Automated Leukemia Detection and Classification

Abstract

Leukaemia is a serious haematological cancer, where the abnormal growth of white blood cells needs to be diagnosed early and accurately, to ensure best treatment and survival. The manual microscopic analysis of blood smear images is a time consuming, labour intensive and skill intensive process which can only be performed by expert haematologists, which has encouraged the development of automated computer-aided blood smear diagnostic systems. This study aims to design an Explainable Hybrid Swin Transformer (XHSA-DCNet) for automatic detection and classification of leukemia from microscopic blood smear images, which combines explainable hybrid Swin Transformer and attention-guided dense convolution network. The suggested framework combines a Dense Convolutional Feature Extraction Module for extracting the fine-grained morphology of each cell and a Swin Vision Transformer for learning global contextual relationships. To effectively learn to fuse the local and global feature representations, a Multi-Scale Cross-Attention Fusion (MSCAF) module is introduced, and an Explainable Gradient-weighted Attention Mapping (X-GAM) mechanism to enhance model interpretability by highlighting diagnostically important regions of leukocytes. In addition, using advanced data augmentation and focal loss optimization for improving generalization and class imbalance problems. Experimental results show its effectiveness in terms of classification accuracy (99.1%), precision (98.9%), recall (99.0%), F1-score (98.95%) and AUC (0.995) outperforming CNN, ResNet50, DenseNet121, and Vision Transformer-based ones. The findings show that XHSA-DCNet is a promising clinical decision support system for haematologists and health care professionals, as it is highly accurate, robust and clinically interpretable for the diagnosis of leukemia.