LK-CIL: Language-guided Knowledge-enhanced Class-Incremental Learning with Pretrained Models for Chest X-Ray Imaging

Chest X-ray (CXR) analysis models deployed in clinical environments must continuously adapt to newly defined or emerging disease categories, while historical patient images are often unavailable due to privacy, storage, and regulatory constraints. This poses a practical challenge for class-incremental learning (CIL): models should learn new classes without replaying old data while preserving previously acquired diagnostic knowledge. Existing replay-based CIL methods are difficult to apply in privacy-sensitive medical scenarios, whereas vision-only replay-free methods lack stable semantic anchors and are prone to representation drift and catastrophic forgetting. To address this challenge, we propose language-guided knowledge-enhanced class-incremental learning (LK-CIL), a replay-free vision-language framework for medical image classification. LK-CIL freezes a pretrained vision transformer and introduces lightweight task adapters for parameter-efficient incremental updates. To provide stable semantic guidance, it constructs class textual prototypes from clinically reviewed LLM-generated radiographic descriptions. These prototypes guide knowledge-aware adapter merging across tasks, aligning evolving visual representations with class-level linguistic priors. During inference, LK-CIL refines class prototypes using a graph convolutional network on a class-relation graph, selects the most suitable adapter based on feature-prototype similarity, and applies a self-refined prediction strategy to improve robustness. Experiments on ChestX-ray14 show that LK-CIL achieves a favorable balance between recognition accuracy and forgetting suppression compared with state-of-the-art (SOTA) CIL methods under multiple incremental and distribution settings. Additional evaluations further support its effectiveness on an emerging COVID-19 class, multi-label CXR classification, external CXR datasets, and other medical imaging modalities, including colon histopathology and dermoscopy. These results suggest that LK-CIL provides a practical replay-free solution for continuously updating medical image classification models without storing historical patient data. Code is available at: https://github.com/Lingling-Yuan/LK-CIL.