Mitigating Domain Shift in Concept Bottleneck Models for Diabetic Retinopathy Grading via Learnable Concept Projection

Concept bottleneck models (CBMs) offer interpretable diabetic retinopathy (DR) grading by routing predictions through clinically meaningful lesion concepts, yet they assume that training and deployment data share the same distribution. In practice, fundus images vary substantially across hospital sites due to differences in camera hardware, field of view, and patient demographics, causing concept predictors to misfire on out-of-distribution inputs. To address this limitation, we propose DICAN, a few-shot domain incremental learning framework that freezes all diagnostic components (backbone, prototype bank, and classification head) and trains only a low-rank residual projector (262K parameters). This projector realigns backbone features with frozen concept prototypes when a new hospital domain arrives with as few as 10 labeled images per grade. A multi-cluster prototype bank with best-match-then-aggregate scoring compresses sub-cluster similarities into three fixed statistics per concept, decoupling prototype granularity from classifier complexity and stabilizing few-shot adaptation. In this study, we evaluate DICAN across four sequential sessions using five multi-national DR datasets (DDR, FGADR, APTOS, Messidor-2, DRAC 2022). The proposed method achieves 54.37% average accuracy and 0.48 quadratic weighted kappa after all sessions, outperforming the strongest baseline by 10 percentage points in accuracy while halving forgetting. The concept bottleneck remains fully inspectable and correctable across all sessions, allowing clinicians to override lesion-concept scores and observe deterministic changes in the diagnostic output.