ADAPT-Net: Entropy-guided dynamic architecture expansion with elastic weight consolidation for continual medical imaging

Medical imaging AI systems face two critical challenges: catastrophic forgetting, wherein models lose previously acquired diagnostic knowledge when trained on new data, and distribution shift, wherein evolving clinical conditions degrade model performance over time. This study proposes ADAPT-Net (Adaptive Diagnostic AI for Precision-Tuned Networks), a continual learning framework built on DenseNet-121 that addresses both challenges through three co-designed mechanisms: Elastic Weight Consolidation (EWC) with Fisher information matrices to protect critical diagnostic parameters; a Fisher-weighted experience replay buffer prioritizing rare and uncertain cases for memory-efficient retention; and entropy-guided dynamic architecture expansion that increases capacity only when prediction uncertainty exceeds a novelty threshold (τ = 1.3 bits). The framework is evaluated on the CheXpert dataset (224,316 chest radiographs, 65,240 patients) across five thoracic pathologies Pneumonia, Cardiomegaly, Edema, Pleural Effusion, and Atelectasis under a three-phase task-incremental protocol. ADAPT-Net achieves 7.11% higher diagnostic accuracy, 43% faster inference, 49.17% smaller model size, and 45.23% lower prediction entropy compared to MetaRL, MDCNN, and GPT-4 + DCNN baselines, while reducing catastrophic forgetting from 0.27 to 0.05. These results confirm ADAPT-Net as an accurate, efficient, and adaptable diagnostic framework for diverse healthcare environments.