Meta-attention fusion and adaptive optimization for explainable diagnosis of acute lymphoblastic leukemia in blood smears

The handling of acute lymphoblastic leukemia (ALL) which is the most common malignancy in children, depends on a prompt and precise diagnosis. Conventional techniques, such as bone marrow aspiration followed by flow cytometry, provide reliable results but have serious limitations. These include being invasive, requiring skilled handling, consuming valuable time, and occasionally yielding unclear readings. We present ConfuNetHMA, a novel automated diagnostic tool that analyses peripheral blood smear pictures in order to get over these obstacles. Fundamentally, our approach combines two complementing deep learning architectures: vision transformers to understand broad correlations throughout the entire slide, and convolutional neural networks to capture fine-grained, localized patterns in individual cells. By combining local and global information into a single, coherent picture, these elements communicate via a layered “meta-attention” mechanism. A lightweight meta-learning-inspired module produces an optimiser-selection policy conditioned on the fused feature representation; in this work it is used to identify and recommend the best-performing optimiser per dataset across runs. To confirm its robustness, ConfuNet-HMA was tested both on a fixed train/validation/test split and via five-fold cross-validation. The model reaches 100% classification accuracy on ALL dataset and 97.39% accuracy on C-NMC dataset and maintains an ROC-AUC of 0.994 in differentiating ALL subtypes, with uniformly high precision and recall across categories. Recognizing that clinicians need to trust AI decisions, we’ve integrated several explainability tools. Grad-CAM highlights which regions of each cell influenced the diagnosis, while LIME, SHAP, and Integrated Gradients offer complementary views into feature importance pinpointing nuclear shapes, chromatin textures, and cytoplasmic details that pathologists routinely examine. Because these highlighted features align with established haematological criteria, practitioners can follow the AI’s “thought process” rather than treating it as an inscrutable black box.