Shape Prior for Quality Assessment in OCTA via Denoising Autoencoders at the Segmentation Level

Accurate medical image segmentation is a critical step for detecting and quantifying diseases, identifying pathological regions, and guiding treatment decisions. However, ensuring the reliability of these segmentations, especially when dealing with images from different devices, acquisition protocols, or calibration settings, is essential for clinical trust and safety. We propose a Shape Prior Quality Assessment (SPQA) Framework, which combines uncertainty estimation, denoising autoencoding at the segmentation level as a shape prior, and statistical features to train a Random Forest model for reliable quality prediction. While the segmentation model itself is trained, the approach avoids end-to-end learning on image data, enhancing robustness and interpretability. Experimental results on three external datasets show that our method outperforms existing baselines and remains robust to domain shifts. This framework supports scalable and generalizable deployment of medical AI systems in real-world settings where manual annotations are limited or not feasible. Code will be made publicly available upon publication.