QPSO-LF-MOBO: A Quantum-Inspired Evolutionary Deep Learning Framework for ALL Detection and Localization in Digital Pathology Images

Abstract

Acute Lymphoblastic Leukemia (ALL) diagnosis is hindered by high inter-observer variability (up to 23%), long processing times (6–8 hours), and limited clinical reproducibility. This paper proposes QPSO-LF-MOBO, a quantum-inspired evolutionary neural architecture integrating Capsule Networks, Quantum-Inspired Particle Swarm Optimization with Lévy Flight (QPSO-LF), Multi-Objective Bayesian Optimization (MOBO) with Expected Hypervolume Improvement (EHVI), and Swin Transformer-based cross-attention for joint hierarchical classification and precise localization of leukemic blast cells in digital pathology images. The framework simultaneously optimizes accuracy, inference speed, and model complexity through a unified Pareto-optimal pipeline, further enhanced by contrastive learning and Monte Carlo Dropout-based uncertainty quantification for reliable clinical decision support. Evaluated on four benchmark datasets (ALL-IDB1, ALL-IDB2, C-NMC, and LISC), QPSO-LF-MOBO achieves 99.2% accuracy, 98.9% F1-score, 99.3% specificity, and an IoU of 0.972, with a 3.4× FLOPS improvement over 15 state-of-the-art methods while operating at real-time speed (42.3 FPS), making it a robust and deployable solution for clinical ALL diagnosis.