Uncertainty-Aware Quality Control for Multi-Structure 2D Echocardiography Segmentation

Abstract

This project evaluates entropy, MC Dropout, test-time augmentation, and ensemble disagreement to detect multi-structure segmentation errors and boundary failures in 2D echocardiography. By combining spatial uncertainty with anatomical validity checks, it predicts downstream LV volume and EF errors to build an automated high-, medium-, or low-reliability QC trigger for clinical workflows.