Uncertainty aware pectoral muscle segmentation based on heteroscedastic regression

Paul Zech, Christian Huemmer, Christopher Syben, Ludwig Ritschl, Sebastian Stober

Keywords: Pectoral muscle segmentation, uncertainty, heteroscedastic regression

TL;DR: We present a novel method for pectoral muscle segmentation in mammograms that models heteroscedastic uncertainty.

Abstract:

Accurate segmentation of the pectoral muscle is crucial for improving breast cancer diagnosis in mammograms. While modern deep learning models excel in segmentation, they often lack uncertainty quantification, which is essential for reliable clinical decisions. In this
work, we propose a novel method for modeling uncertainty in pectoral muscle segmentation by combining the prediction of probabilistic heatmaps with heteroscedastic regression. For that, we investigate both an existing and a novel loss function derived from the heteroscedastic Laplace distribution, and show that our loss function is more robust in our setting for pectoral muscle segmentation. Further, we demonstrate that our method is capable of producing heatmaps with high-likelihood predictive distributions within a single model while outperforming an ensemble baseline in terms of accuracy.

Submission Number: 54