EVA Net: Adaptive Attention Gating with Parallel Dilated Convolution for Multi Scale CNN ViT Fusion in Leukemia Classification

Abstract

The early and accurate diagnosis of leukemia is necessary for effective treatment. Traditional methods of diagnosis, including manual microscopy of peripheral blood smears, are subject, time-consuming and produce inconsistent results. The use of advanced methods such as hybrid CNN-RNN models and CNN with Vision Transformers (ViT) suffers from difficulties in capturing fine-grained interactions and identifying subtle morphological differences, particularly in leukemic cells. This paper introduces a dual backbone network (EVA-Net) designed to improve the classification performance. EVA-Net introduces Parallel Dilated Convolution (PDC) in the Multi-Scale Refinement Block (MSRB) to increase the receptive field of the analysed features and Adaptive Attention Gating (AAG) for the spatially-adaptive integration of EfficientNet local features and ViT global context. In addition, it adopts a memory-saving training strategy, incorporating gradient checkpointing, mixed precision and gradient accumulation, which allows it to be deployed in standard GPUs as well. The architecture enables the building of a powerful, efficient multi-class classification model for leukemia that outperforms the current ones, and that can be made interpretable via attention visualisation. EVA-Net was able to achieve an accuracy of 99.59%, a macro F1 score of 0.9959 and a Matthews correlation coefficient of 0.9944 on 3,256 images of peripheral blood smears across four classes, significantly outperforming competitive models like ResNet50 (93.05%) and EfficientNetB3 (97.14%). McNemar's test (p<0.01) confirmed that statistical significance was extremely high and the relative error was found to be 31.5% greater than the best one obtained in ablation studies.