When does domain-specific pre-training help? A controlled comparison of retinal foundation models for OCT classification — reproducible code

Reproducible analysis code for a controlled comparison of four frozen image encoders (ImageNet-ViT, MAE-ViT-L, RETFound-fundus, RETFound-OCT) as feature extractors for seven-class OCT classification on the public OCTDL dataset, with patient-level splitting, calibration analysis, bootstrap confidence intervals, and McNemar tests. Accompanies the manuscript submitted to Graefe's Archive for Clinical and Experimental Ophthalmology.