CADRE: Stable, Parameter Efficient Adaptation of Medical Vision Language Models with Bounded Forgetting and Prior Drift

The article presents CADRE, a framework for the continual adaptation of medical vision-language models (VLMs) like BiomedCLIP, focusing on mitigating catastrophic forgetting and prior drift during updates for new imaging modalities. CADRE employs a frozen-backbone architecture that integrates low-rank adaptation (LoRA) with a similarity-aware elastic weight consolidation mechanism, achieving a reduction in forgetting by approximately sevenfold and demonstrating superior accuracy and backward transfer across diverse modalities (histopathology, ultrasound, chest radiography) while only training about 0.23% of the parameters. This work is significant for practitioners as it addresses critical safety and reliability concerns in deploying adaptive models in clinical settings, emphasizing stability over traditional accuracy benchmarks.