Stage-Specific Image Enhancement for Diabetic Retinopathy Image Classification

Abstract

Screening for diabetic retinopathy (DR) depends on pipelines that are transparent and reproducible, yet image preprocessing is often under-specified or coupled with model changes, obscuring attribution. This work asks whether stage-aware preprocessing - targeted, interpretable edits aimed at stage-defining cues - can shift five-class DR grading when everything else is held constant. Using APTOS-2019 with fixed patient-level splits, we establish a neutral baseline that standardizes geometry and framing, select a single modern pretrained backbone as the reference, and then freeze the training recipe. We vary only the input preparation through five scenarios (S0–S4) aligned with ICDR stages 0-4, applying each uniformly to all images to keep comparisons fair. Each scenario yields a dataset variant for training a single model under identical settings, followed by evaluation on the unchanged test set with standard multi-class reporting to reveal class-specific shifts and adjacent confusions. Across scenarios, targeted edits produce stage-specific gains without destabilizing overall behavior: stabilization supports a balanced profile, proliferative-focused edits enhance advanced-stage detection, and vessel-emphasis benefits severe NPDR, while dot/bright-centric variants largely track the baseline on this split. The study is bounded by a single public corpus, class imbalance, and image-level labels; consequently, findings should be read as dataset-internal evidence that input design matters. Future work should assess cross-dataset robustness, incorporate lesion-level supervision or saliency checks to verify where edits act, and pair stage-aware inputs with calibration and interpretability suitable for screening deployment.