FAZ Segmentation Quality Assessment in OCTA via Denoising Autoencoders and Segmentation Uncertainty Estimation
Keywords: Quality Assessment, Uncertainty, Segmentation, Image Domain Shift
TL;DR: A Segmentation Quality Assessment Framework method that estimates segmentation quality without relying on ground-truth labels.
Abstract: Accurate segmentation quality assessment is essential in medical imaging, particularly in preventing segmentation algorithms from failing silently. We propose a Segmentation Quality Assessment Framework method that estimates segmentation quality without relying on ground-truth labels. Our approach integrates learning-free uncertainty estimation with a Denoising Autoencoder (DAE) to generate pseudo-labels, extract key statistical features, and train a Random Forest Regressor (RDF) for quality prediction. Experimental results demonstrate that our method outperforms baseline approaches on three external datasets, showcasing its robustness to image domain shifts. our method enhances the scalability and generalizability of real-world medical imaging applications by leveraging segmentation models to handle cases where manual annotations are missing or infeasible.
Submission Number: 14
Loading