Alireza Norouziazad, Fatemeh Esmaeildoost, Razieh Salahandish

Abstract: Accurate segmentation of disease biomarkers, such as hippocampal atrophy in Alzheimer’s disease (AD) and tumor regions in breast ultrasound, is essential for early diagnosis and treatment planning. To address feature redundancy and poor interpretability in deep segmentation models, we propose a dynamic, training-time, fuzzy-adaptive channel selection framework guided by moth-flame (MFO) and whale (WOA) optimization algorithms. Our method enhances multi-scale feature representation by fitness-driven selection of discriminative channels without modifying network architecture or requiring retraining. Evaluated on clinically curated AD MRI and breast ultrasound datasets, our approach consistently improves segmentation performance: on hippocampal segmentation, it achieves 99.7% sensitivity and 85.4% Dice score, surpassing baselines by 0.6% and 5.3%, respectively. The fitness-guided mechanism also provides interpretable feature selection, aligning with clinical demands for trustworthy AI.

Keywords: Medical image segmentation, Alzheimer’s disease, breast cancer, metaheuristic optimization, fuzzy logic, channel selection, U-Net, DeepLabV3+, interpretability.

Date Published: December 9, 2025
DOI: 10.11159/jmids.2025.007

View Article