Uncertainty Estimation in Semantic Segmentation of Ultrasound Images
This project increases reliability of automated ultrasound interpretation and supports task-shifting in medical diagnostics, by developing uncertainty quantification methods for automated ultrasound image segmentation.
2D Ultrasound (US) imaging is crucial for diagnosing cardiovascular diseases and conducting fetal scans due to its portability and low cost. However, its operator-dependent nature restricts usage to locations with expert personnel. While Deep Learning (DL) shows potential for automation, the absence of uncertainty modeling in current DL methods limits real-time feedback during image acquisition, affects image quality, and leads to inaccuracies during automated measurement and interpretation. The performance of uncertainty models can also vary significantly.
Our goal is to develop novel methods for uncertainty modeling in semantic image segmentation from US images. By addressing the gap in uncertainty estimation, we aim to enhance the reliability and accuracy of automated interpretation of US images.
This project successfully developed methods for uncertainty estimation in semantic segmentation of ultrasound images. The techniques improve the reliability and accuracy of automated interpretation for cardiovascular and fetal scans by quantifying model confidence and identifying potential segmentation errors. This work lays the groundwork for more robust, operator-independent diagnostic tools that can provide real-time feedback during image acquisition and support task-shifting of ultrasound diagnostics to non-expert operators.