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

Abstract

Segmentation models deployed in clinical practice need to meet quality standards for each image, and even when models perform well on average, they may fail to segment individual images with sufficiently high quality. We propose a combined quality control and error correction framework that recommends the number of local patches requiring manual review and estimates the resulting impact on the Dice Score of the corrected segmentation, allowing a trade-off between segmentation quality and time invested in manual review. Patches are selected for correction based on uncertainty maps obtained from an ensemble of segmentation models, demonstrated on retinal vessel segmentation.

Jeremiah Fadugba (Jerofad)

Data Scientist | Machine Learning Engineer

My current research interests include Medical Imaging, Machine Learning, Deep Learning, and Trustworthy ML.