A retinal-image classifier built around rings of blood-vessel measurements matched RETFound on HRF, but lagged it on FIVES and SUSTech-SYSU. With predicted vessel masks, the ring-based method reached 91.1% accuracy on HRF, equal to RETFound's score. It reached 72.4% on FIVES, below RETFound's 80.6%, and 89.6% on SUSTech-SYSU, below RETFound's 94.6% and ConvNeXt-B's 96.4%. The reported 95% accuracy intervals for the ring method ran from 79.3% to 96.5% on HRF, 67.7% to 76.6% on FIVES and 87.5% to 91.3% on SUSTech-SYSU. With expert vessel masks, its accuracy was 100% on HRF and 76.6% on FIVES.
A classifier built around the optic disc
The method was designed to make the image cues easier to inspect. It places concentric rings around the optic disc and extracts four families of vascular descriptors from those zones. The descriptors combine vessel geometry, red-green color, an RGB-derived SO2 proxy and vessel-background entropy, a measure of variation in the surrounding image. The combined measurements are standardized and passed to an elastic-net logistic regression classifier. The study's central comparison was between this explicit vascular information and learned deep image representations.
Three public image collections
Three public image collections supplied the test data: HRF, selected high-quality FIVES images and SUSTech-SYSU. HRF contained 45 images, with 15 healthy, 15 showing diabetic retinopathy and 15 showing glaucoma. The selected FIVES set contained 168 healthy images, 83 glaucoma images and 130 diabetic-retinopathy images. SUSTech-SYSU contained 1,016 images, including 631 healthy and 385 diabetic-retinopathy images. HRF and FIVES were assessed with leave-one-out cross-validation, in which each image is held out in turn, while SUSTech-SYSU used stratified five-fold cross-validation. Scores were calculated from aggregated out-of-fold predictions.
The vessel map still matters
The vessel masks were part of the test, not just a technical detail. When the researchers changed the segmentation algorithm, RIP-AV produced 91.1% accuracy on HRF. On FIVES, it produced 72.4%, almost the same as SA-UNetv2 at 72.2% and above FR-UNet at 68.5%. The reported results showed larger algorithm differences on HRF and closer results on SUSTech-SYSU.
Image framing changed the deep models
FIVES also exposed how much image framing could matter to whole-image models. After the field of view was standardized, RETFound's accuracy fell from 92.7% to 80.6%. ConvNeXt-B fell from 82.9% to 76.9%, ViT-B/16 from 81.9% to 75.3% and ResNet-50 from 79.8% to 72.4%. The experiment links the models' scores to spatial characteristics of the images in this setting, but does not isolate acquisition-related variation from clinically meaningful background abnormalities.
Changing the non-vessel background produced another drop for RETFound. Starting from the field-of-view-standardized accuracy of 80.6%, the score fell to 72.5% after background color standardization, 76.3% after contrast reduction, 76.0% after brightness standardization and 68.2% after all three changes together. By comparison, lesion-aware inpainting changed overall accuracy from 80.6% to 80.3% and diabetic-retinopathy recall from 83.8% to 83.1%. That last test does not establish independence from lesion information: incomplete lesion detection and residual contextual or textural information may have remained.
A color-based feature carried information
Within HRF, removing the SO2-proxy features was associated with the largest observed ablation change: accuracy fell from 91.1% to 77.8%. Removing vessel-background entropy also reduced performance. In a separate HRF test using expert vessel masks, classification generally improved as more peripheral rings were added, particularly for combined and geometry-based feature sets. The findings point to information spread across the representation, but the SO2 proxy should not be read as a calibrated measurement of retinal oxygen saturation.
Transfer was the hardest test
The one-way FIVES-to-HRF transfer test was harsher for the compared image baselines. ResNet-50 and ConvNeXt-B each reached 35.6% accuracy, while ViT-B/16 and RETFound each reached 33.3%. The ring representation reached 73.3% without RGB color and SO2-proxy features, and 77.8% with a normalized SO2-proxy descriptor. The reported 95% accuracy intervals were 59.0% to 84.0% and 63.7% to 87.5%, respectively. In this specific transfer experiment, the ring representation performed better than the compared baselines, but that does not establish general superiority across institutions, devices or populations.
A benchmark, not a clinical verdict
The document is a preprint, arXiv:2608.24723v1, dated 25 Aug 2026. Its evidence comes from public-dataset cross-validation, controlled image perturbations, feature ablations and one FIVES-to-HRF transfer direction. This makes it a computational methods benchmark, not a prospective clinical validation study or evidence of improved patient outcomes. The reported framework also depends on vessel segmentation, optic-disc localization and stable color and intensity measurements.
Paper data and sources
Original title: Interpretable Fundus Image Classification via Ring-Based Retinal Vasculature Features
Authors: Xiaoyan Li, Shixin Xu, Arvind Gupta, Huaxiong Huang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text