Deep Learning–Based Dual‐Pipeline Framework for Acute Lymphoblastic Leukemia Classification With Explainable AI Integration

Acute lymphoblastic leukemia (ALL) is a type of blood cancer that rapidly impacts the hematologic system, particularly in children, requiring accurate and prompt diagnosis for better clinical outcomes. Manual inspection of peripheral blood smears remains the standard approach, but it is resource‐intensive and subject to observer variability, underscoring the need for automated, consistent, and interpretable diagnostic tools. This work proposes a dual‐stage deep learning framework combining transfer learning and a custom CNN model—DAC‐Net—for the classification of ALL from microscopic blood smear images. The first stage leverages pretrained CNNs (VGG19, ResNet50, ResNet101) for feature extraction, followed by feature refinement using ANOVA, recursive feature elimination (RFE), and random forest importance scores. These features are classified using multiple traditional machine learning algorithms including SVM, kNN, random forest, and Naïve Bayes. The second pipeline features the DAC‐Net, an attention‐integrated CNN trained end‐to‐end to automatically learn discriminative patterns directly from image data. Incorporating spatial attention layers and dense connections, the DAC‐Net emphasizes morphologically relevant regions associated with leukemia. Extensive experiments on the C‐NMC dataset demonstrated that while both pipelines performed well, the DAC‐Net achieved higher sensitivity and F1‐scores. The integration of Grad‐CAM and related explainability tools adds transparency to model decisions, enhancing its practical value as a decision‐support tool in ALL diagnosis.