Do Global Vessel Descriptors Help Multi-Disease Fundus Classification? A Leakage-Controlled Ablation of Hybrid CNN–Transformer Models with External Testing

Background/Objectives: Retinal vessel features are often added to fundus classifiers without isolating their effect. We examined whether global vessel descriptors improve multi-disease fundus classification under leakage-controlled external testing. Methods: After a duplicate audit and field-of-view normalization, models were trained on RFMiD and APTOS with a partial-label mask and tested internally and on four external datasets (ODIR-5K, Mendeley, STARE, IDRiD) that were not used for training, threshold fitting or calibration. Fourteen arms were compared: nine architectures combining EfficientNet-B4, Swin-S and a vessel branch, a label-handling control, and four variants with vessel features from other segmenters; ten arms were trained with three seeds. Vessel descriptors came from segmenters trained on 14 DRIVE images, 14 FIVES images or 480 FIVES images. Paired differences were tested with a hierarchical bootstrap and Holm correction; all external comparisons are exploratory. Results: The CNN–Transformer hybrid reached an internal macro-AUROC of 0.945 ± 0.002 and a mean external macro-AUROC of 0.830 ± 0.005, exceeding the CNN baseline internally, on ODIR-5K and on Mendeley (Holm-adjusted p ≤ 0.0045) but not on STARE or IDRiD; an average of separately trained CNN and Transformer models performed similarly (0.831). Treating unassessed labels as negatives lowered ODIR-5K and Mendeley AUROC (−0.018 and −0.032). For the CNN vessel arm, FIVES test Dice of 0.662, 0.779 and 0.833 for the three segmenters was accompanied by mean external AUROC of 0.786, 0.792 and 0.805, with the clearest differences on STARE and Mendeley; no vessel arm exceeded its counterpart without vessels. Conclusions: Global vessel descriptors added no measurable value to the image backbones studied, and their effect was associated with the upstream segmentation pipeline, whose quality should therefore be reported.