A framework for leukemia detection based on multi fusion strategy

Accurate and early detection of Acute Lymphoblastic Leukemia (ALL) is critical for improving patient outcomes, yet the adoption of deep learning based diagnostic systems in clinical practice remains limited due to their black-box nature. In this article, we proposed an explainable deep learning (DL) framework for ALL detection from microscopic blood smear images that combines high diagnostic performance with transparent decision-making. The framework employs an EfficientNetB4-based classifier enhanced with multiple Explainable Artificial Intelligence (XAI) techniques, including Local Interpretable Model-agnostic Explanations (LIME), Gradient-weighted Class Activation Mapping (Grad-CAM), and Guided Gradient-weighted Class Activation Mapping (Guided Grad-CAM). To address the limitations of individual explanation methods, a novel Combined XAI strategy is introduced, leveraging multiple fusion mechanisms to generate consolidated, reliable visual explanations. The proposed model is evaluated on a publicly available ALL dataset and achieved an accuracy of 0.92, F1-score of 0.89, Cohen’s Kappa of 0.87, and MCC of 0.85 on the independent test set, outperforming several well-known transfer learning baselines. In addition to classification performance, explanation quality was assessed using stability and deletion metrics, where the Combined XAI approach showed the best robustness and faithfulness, with a stability score of 0.680922 and a positive deletion score of 0.231156. Expert pathologists also confirmed that the generated explanations aligned well with clinically relevant regions.