Bridging the Ophthalmic Screening Gap

Evaluating Diffusion-Based and Classical Preprocessing for Low-Resource Eye Disease Classification

Abstract

Automated retinal disease screening in low-resource settings is constrained by the image quality limitations of low-cost fundus cameras, yet the impact of realistic degradation on multi-class classification performance remains poorly characterized. This study investigates whether DDPM-based restoration can stabilize retinal disease classification from degraded fundus photographs. Using a four-class fundus dataset (4,217 images: normal, cataract, diabetic retinopathy, glaucoma), we trained an EfficientNet-B0 classifier under four conditions — high-quality (HQ), synthetically degraded low-quality (LQ), LQ+CLAHE, and LQ+DDPM — evaluated via three-fold stratified cross-validation. Synthetic degradation reduced macro-F1 by 6.0 percentage points and macro-AUC by 1.8 points relative to HQ. Among preprocessing strategies, DDPM-based restoration achieved macro-F1 of 0.8807 and macro-AUC of 0.9780, outperforming both the LQ baseline (F1 0.8587) and CLAHE (F1 0.8604), while reducing cross-fold F1 standard deviation from 0.0174 to 0.0056 — a three-fold improvement in generalization stability. Qualitative analysis further confirmed that DDPM enhances vessel visibility and optic disc clarity without the noise overamplification occasionally introduced by CLAHE. These findings position DDPM-based preprocessing as a compelling strategy for robust retinal disease classification in resource-constrained screening environments.