EmAlexNet: Deep Multi-Scale Modified AlexNet Ensemble With Gated Residual Fusion and Dual Attention for Acute Lymphoblastic Leukemia Classification

Abstract

Accurate classification of acute lymphoblastic leukemia (ALL) blood-smear images remains challenging because diagnostically relevant morphology occurs at multiple spatial scales. High-performing deep models may also exhibit correlated errors and poorly characterized prediction confidence. This work proposes ensemble modified AlexNet (EmAlexNet), a deep multi-scale ensemble composed of three independently trained modified AlexNet (mAlexNet) learners. The proposed architecture combines multi-scale residual feature extraction, sigmoid-gated cross-stage fusion, residual channel recalibration, spatial self-attention, and mean–maximum feature aggregation. The constituent learners differ in random initialization and training objective, using cross-entropy, focal loss, and Kullback–Leibler (KL) divergence to a fixed label-smoothed target. Their class probabilities are combined using fixed equal-weight averaging, without an additional meta-classifier. The method was evaluated on the 20,000-image ALL subset of the Kaggle Multi Cancer Dataset. The dataset contains benign hematogone, early pre-B, pre-B, and pro-B categories. Under stratified image-level 10-fold cross-validation, EmAlexNet achieved 99.55% accuracy, 99.55% macro F-score, and a macro one-versus-rest area under the receiver operating characteristic curve (AUC) of 0.9981. Probability averaging also reduced residual classification errors and improved calibration relative to the strongest individual mAlexNet learner. These findings support the utility of combining matched architectures trained with different objective formulations and random initializations under the evaluated repository-level protocol. However, patient, specimen, and parent-image identifiers were unavailable. Therefore, patient-independent generalization cannot be established. External patient-grouped and multicenter validation is required before the clinical utility of EmAlexNet can be assessed.