| Abstract [eng] |
The retinal microvasculature provides valuable biomarkers for systemic diseases, yet accurate segmentation of thin retinal vessels remains challenging, particularly under resolution constraints imposed by modern deep learning pipelines. While many retinal vessel segmentation studies report strong overall performance, they rarely distinguish between vessel calibers or quantify how image and mask downscaling disproportionately affects thin vessels that carry important clinical information. This work systematically investigates how fundus image and vessel mask resolution reduction impacts thin retinal vessel structure and subsequent segmentation accuracy. We establish a baseline retinal vessel segmentation model using high-resolution images from the FIVES dataset and the HRFormer architecture and propose a morphological algorithm to separate thin and thick vessels in binary segmentation masks based on vessel thickness and image resolution. This enables class-specific evaluation without reframing segmentation as a multi-class problem. Using this framework, we quantify structural information loss (defined as F1-score error) introduced in ground truth masks by different mask downscaling strategies. We also assess segmentation performance across progressively lower fundus image resolutions while keeping model architecture and training conditions fixed. Our results show that linear interpolation methods (bilinear, bicubic and distance transform) best preserve vessel structural information during mask downscaling, but require resolution-specific threshold calibration, as the conventional 50% intensity threshold is suboptimal across scales. Thin ground truth vessels are substantially more affected by resolution reduction than thick vessels: even under optimal conditions, thin vessel F1-score decreases from 0.9286 at 1024x1024 to 0.7014 at 256x256, that is 2.6 times more structural information loss than thick vessels for given resolutions. Overall, reducing mask resolution by a factor of 2 constitutes an equivalent drop in structural information for both thin and thick vessels which leaves use of downscaled resolutions in retinal research a questionable choice. Fundus image resolution experiments reveal that 512x512 constitutes a practical lower bound for thin vessel analysis, as models trained at 1024x1024 and 512x512 achieve statistically indistinguishable performance, though a big disparity is left between thin and thick vessels (F1-score of 0.70 for thin and 0.91 for thick vessel segmentation). Further reducing the resolution significantly affects thin vessel predictions while thick vessels remain relatively intact. Recall is identified as the primary metric impacted by resolution reduction, indicating loss of vessel connectivity rather than boundary distortion. Despite comprising only about 15% of vessel pixels, thin vessels contribute approximately 78% of normalized segmentation errors at practical resolutions, highlighting their disproportionate influence on model output quality. These findings demonstrate that overall segmentation metrics can obscure clinically relevant segmentation failures. |
