Fusing parametric and spatial constraints in continual learning

The deployment of powerful foundation models in medical imaging is severely hampered by the fact that neural networks cannot learn sequentially from decentralised data with- out catastrophic forgetting. Early continual learning frameworks, such as Elastic Weight Consolidation (EWC), offer a parametric defence against forgetting but treat the network as a black box and neglect to preserve learned spatial representations. To solve this problem, we propose a novel continual learning framework that anchors EWC within a stable reconstruction-classification training paradigm using batch-trained prototypes. Our approach fundamentally constrains representation drift by enforcing spa- tial alignment between the network’s dynamically generated saliency maps and static visual class prototypes. Our approach combines weight consolidation with spatial reconstruction penal- ties to explicitly prevent the prior classification state-space from being distorted during novel task updates. Empirical evaluations show that the fusion not only retains high classification efficacy on historical tasks but also successfully maintains the network’s visual interpretability. Ultimately, this framework establishes a new direction for continual learning in medical imaging, shifting the focus from purely weight-based regularisation to the holistic preservation of both classifier stability and interpretable feature representations.