Task-Specific Quality Gating for Retinal Optical Coherence Tomography B-Scans: Learned Representations Over Scalar Metrics in Choroid Segmentation

Device-reported scalar image-quality scores are commonly used to gate deep-learning segmentation of retinal OCT B-scans, but such aggregate statistics can miss localized failure modes relevant to specific anatomical targets. We evaluate vision foundation models as learned-representation proxies for failure prediction in task-specific (e.g., choroid layer) segmentation quality gating, and find that RetFound, an OCT-specific foundation model, outperforms both general-purpose vision foundation models and traditional no-reference image-quality assessment (NR-IQA) methods, which underperform due to their reliance on aggregated image statistics. The proposed approach reduces silent failures in clinical deployment. A. Gadari is the corresponding author.

In Proceedings of IEEE Healthcom 2026, New York, NY, USA. Paper