HemaNest-UNet++ and explainable vision transformer for acute lymphoblastic leukemia classification

Acute Lymphoblastic Leukemia (ALL) is a rapidly progressing hematological malignancy in which early and reliable identification of abnormal lymphoblasts is essential for treatment planning and prognosis. Although deep learning has improved automated leukemia analysis, most existing studies focus on binary classification, rely on public peripheral-blood datasets, classify entire microscopic images without isolating diagnostically relevant cells, or provide limited interpretation of model decisions. These limitations reduce their suitability for fine-grained and clinically transparent analysis of bone marrow images. To address this gap, this study proposes a segmentation-guided and explainable deep learning framework for automated ALL detection and morphological subtype classification. A custom dataset of 387 expert-labelled bone marrow microscopic images was collected from Amreek Clinical Laboratory, Swat, Pakistan, comprising 117 ALL L1, 103 ALL L2, 62 ALL L3, and 105 Normal samples. Written informed consent was obtained, and all data were anonymized before analysis. The proposed framework first improves image quality through normalization, denoising, and contrast enhancement. It then applies the proposed HemaNest-UNet++, which combines nested skip pathways, dual-attention refinement, multi-scale contextual aggregation, and deep supervision to segment white blood cell regions. The resulting regions of interest are classified using an Explainable Vision Transformer that integrates attention rollout and gradient-based visual explanations. HemaNest-UNet + + achieved a Dice coefficient of 0.934, an intersection-over-union of 0.876, and a pixel accuracy of 0.978, outperforming established segmentation baselines. The proposed classifier achieved 94.8% accuracy, a macro F1-score of 0.944, and a macro AUC of 0.982, exceeding CNN- and transformer-based comparison models. Most residual errors occurred between ALL L2 and ALL L3 because of their overlapping morphology. Explainability maps showed that predictions were primarily influenced by clinically relevant nuclear and cytoplasmic structures. These findings demonstrate the potential of the proposed framework as an accurate and interpretable decision-support approach for bone marrow-based ALL assessment.