Uncertainty Estimation in Semantic Segmentation of US Images

Uncertainty Estimation in Semantic Segmentation of US Images

2D Ultrasound (US) imaging serves as a primary modality for diagnosing cardiovascular diseases and conducting fetal scans due to its portability and relatively low cost. However, the operator-dependent nature of acquiring and interpreting US images restricts its usage to locations with expert personnel. Deep Learning (DL) has exhibited potential in automating view classification and assessing structures and functions from US images, but the absence of uncertainty modeling in these methods poses limitations.


Current DL methods for automated interpretation of US images often lack uncertainty modeling, which is crucial when testing on data distributions different from the training data. The absence of uncertainty estimates can hinder real-time feedback during image acquisition, affecting image quality, and lead to inaccuracies during automated measurement and interpretation. Additionally, the performance of uncertainty models and quantification metrics can vary based on the prediction task and the models under comparison.

Research Aim:

This research aims to develop novel methods for uncertainty modeling in semantic image segmentation from US images. By addressing the gap in uncertainty estimation, the study seeks to enhance the reliability and accuracy of automated interpretation of US images. The focus will be on gaining insights into uncertainty modeling techniques tailored for semantic segmentation tasks, considering the specific challenges and requirements posed by US image data.

Outcome So Far:

Initial investigations have begun into developing novel methods for uncertainty modeling in semantic image segmentation from US images. Efforts are focused on understanding the unique characteristics of US image data and identifying suitable uncertainty estimation techniques. Ongoing research aims to refine these methods and evaluate their performance in providing accurate and reliable automated interpretation of US images, with the ultimate goal of improving diagnostic capabilities in cardiovascular diseases and fetal scans.

Team: Bishesh Khanal, Lavsen Dahal and Aayush Kafle

Relevant Publications

  1. Dahal, L., Kafle, A., & Khanal, B. (2020). Uncertainty Estimation in Deep 2D Echocardiography Segmentation. arXiv preprint arXiv:2005.09349. pdf
Research Themes: Transforming Global health with AI (TOGAI)
Project Category: Computer Vision, Machine Learning, Medical Imaging