Selecting What SSL Sees: Label-Free Top-k Patch Curation for Generalizable Diabetic Retinopathy Grading

Diabetic retinopathy (DR) grading depends on small and unevenly distributed retinal abnormalities, yet self-supervised learning (SSL) commonly receives either resized full images or patches generated by a fixed rule. This paper studies whether the input construction stage can be made more selective without requiring lesion labels or a trainable patch selector. First, two lightweight top-k patch-curation strategies are introduced: a frequency-aware score based on spectral magnitude, and a foreground-aware score combining retinal foreground filtering, sharpness, entropy, and spatial diversity. Second, SimCLR-style contrastive pretraining extended to Vision Transformers (ViTs) to study the effect of input resolution, model scale, pretraining source, labelled-data availability and momentum-contrast pretraining. Experiments use APTOS, EyePACS, IDRiD, Messidor-2, and DDR, followed by frozen linear probing and full-image five-class fine-tuning. ViT-B/384 reaches 0.732/0.789 ACC/QWK without top-k selection and improves to 0.759/0.817 with k = 5 at 100% labelled APTOS. Under adaptive linear probing, foreground aware top-5 sampling reaches 0.516/0.824 ACC/QWK on IDRiD with MoCoV2 (ResNet-50) and 0.527/0.814 with MoCoV3(ViT-B/384). After MoCoV2 fine tuning, frequency-aware top-5 reaches 0.852/0.897 on APTOS, whereas foreground aware top-5 transfers better to several external datasets. The gains are not uniform, showing that patch informativeness, spatial coverage, pretraining source, and domain shift interact.