Evaluating the real-world accessibility of open-source deep learning models for retinal vessel segmentation in optical coherence tomography angiography

Accurate segmentation of retinal vasculature in optical coherence tomography angiography (OCTA) images underpins disease diagnosis and monitoring. While numerous open-source segmentation algorithms are available, an optimal solution remains elusive. This study evaluates and compares the real-world usability and performance of open-source deep learning-based vessel segmentation models across a diverse dataset comprising healthy and diseased eyes. The selected models are standardised in training and performance settings. Model usability is assessed against a range of software engineering criteria defined by the International Organisation for Standardisation. Performance is evaluated on the publicly available OCTA-500 and ROSE-1 datasets using the Dice coefficient of similarity (Dice) and Intersection over Union (IoU). We show that, within the evaluated sample, considerable variability exists in the quality and usability of open-source models, which may limit the practical integration and adoption of some tools within clinical research workflows. Common shortcomings include incomplete or missing dependency documentation, limited portability across systems, hard-coded file paths, and missing scripts or modules. Among the reviewed models, FRNet most closely met the requirements of international software engineering quality standards. In addition to its strong usability profile, FRNet consistently achieved high Dice and IoU across all evaluated datasets. Beyond identifying a reliable OCTA vessel segmentation model, this study highlights a gap between open-source development practices and their practical use in clinical research. While open-source frameworks support reproducibility and translation, explicit usability and quality criteria are needed to ensure real-world adoption.