Abstract
A common hypothesis in retinal screening is that generative AI (GenAI) restoration recovers diagnostic accuracy on degraded fundus images. We test it on APTOS 2019 with nine synthetic degradations over a frozen 440-image test split, two graders (ConvNeXt-Base, CLIP-ViT-B/16) and six restorers spanning CLAHE, GAN super-resolution and diffusion. The result is negative: restoration improves pixel fidelity but not diagnostic performance, and both diffusion restorers we trained actively harm it. A clean-image control isolates the mechanism (passing undegraded images through Cold Diffusion costs ConvNeXt-Base 0.60 accuracy), and a paired analysis shows its shape: restorers systematically move predictions, Cold Diffusion inflating grades and the pathology-preserving DDPM collapsing them onto “No DR”, preserving majority-class accuracy while halving mean recall over the referable grades, so quadratic weighted kappa (QWK) exposes harm that accuracy hides. Attributions also lose stability and perturbation-based faithfulness before accuracy falls. Across most of the risk–coverage curve, quality-aware triage beats do-nothing and restore-all for the CNN (κ_w = 0.665 vs. 0.525 and 0.464 at 80% coverage) but not for the more robust ViT. Sub-threshold images should be re-acquired, not restored.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/OMIA_001.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=DKIDxxCXUH
BibTex
@InProceedings{KocAth_Towards_MICCAISAT2026,
author = { Kocharekar, Atharva AND Painuly, Naveen AND Mileo, Alessandra},
title = { { Towards a Robust and Explainable Pipeline for Diabetic Retinopathy Classification through Quality-Aware GenAI Image Restoration } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17270},
month = {pending},
page = {pending}
}