Abstract

Fundus fluorescein angiography (FFA) is the gold standard for diabetic retinopathy (DR) staging, yet whether angiographic phase affects deep learning performance remains unknown. Using 7508 FFA images from 863 eyes, stratified into venous, recirculation, and late phases, we developed Swin Transformer- and ConvNeXt-based DR staging models under the International five-grade and Chinese six-grade classification systems. Multiclass and binary (NPDR vs. PDR) classification tasks were evaluated. This study provides the first systematic quantification of angiographic phase effects on FFA-based DR staging. Phase-related performance was analyzed using generalized linear mixed-effects models with beta regression. ConvNeXt generally outperformed Swin Transformer, particularly in multiclass classification. Across all experimental settings, performance showed a consistent numerical decline from the venous to the late phase. Under the International five-grade system, ConvNeXt accuracy declined from 86.67% to 81.87%; however, Bonferroni-adjusted comparisons revealed no statistically significant phase-related differences (all adjusted p > 0.05), with small effect sizes (|SMD| < 0.2). Binary classification remained highly stable across phases, with accuracies exceeding 91%. Grad-CAM visualizations demonstrated progressively diffuse model attention in later phases. These findings support phase-flexible FFA acquisition for binary DR screening, whereas venous- or recirculation-phase images remain preferable for high-precision multiclass staging, guiding phase-aware AI development.

IPC Classification

Keywords