Echocardiography is central to measuring clinical parameters like left ventricular volume and ejection fraction, which help diagnose heart conditions. When automated tools compute these parameters, clinicians need to know how much to trust the results. However, standard segmentation-based methods produce pixel-level uncertainty that does not easily translate into uncertainty about the final clinical metric.
The authors propose a contour-based alternative. Instead of segmenting the heart chamber, their model explicitly predicts uncertainty in the contour location, from which multiple plausible contours can be sampled. These sampled contours are then used to propagate uncertainty through to the clinical metric calculation, giving a direct estimate of confidence in the output.
The method was tested on one public and two private cardiac ultrasound datasets, covering 1,264 patients in total. The authors report that it yields accurate uncertainty estimates both for the contouring task itself and for the downstream clinical metrics. Code is available, which should help others build on the approach and compare against it in future work.