Efficiently Correcting Patch-Based Segmentation Errors to Control Image-Level Performance in Retinal Images

Abstract

Segmentation models which are deployed into clinical practice need to meet a quality standard for each image. Even when models perform well on average, they may fail at segmenting individual images with a sufficiently high quality. We propose a combined quality control and error correction framework to reach the desired segmentation quality in each image. Our framework recommends the necessary number of local patches for manual review and estimates the impact of the intervention on the Dice Score of the corrected segmentation. This allows to trade off segmentation quality against time invested into manual review. We select the patches based on uncertainty maps obtained from an ensemble of segmentation models. We evaluated our method on retinal vessel segmentation on fundus images, where the Dice Score increased substantially after reviewing only a few patches. Our method accurately estimated the review’s impact on the Dice Score and we found that our framework controls the quality standard efficiently, i.e. reviewing as little as necessary.