Structure Size as Confounder in Uncertainty Based Segmentation Quality Prediction

Kai Geißler, Jochen G. Hirsch, Stefan Heldmann, Hans Meine

Keywords: Medical Image Segmentation, Uncertainty Quantification

Abstract: Various uncertainty estimation methods have been proposed for deep learning-based image segmentation models. An uncertainty measure is treated useful if it can be used to accurately predict segmentation quality. Therefore, structure-wise uncertainty measures are frequently correlated with measures like the Dice score. However, it is known that the Dice score highly depends on the size of the structure of interest. It is less well-known that popular structure-wise uncertainty measures also correlate with structure size. Therefore, the structure size acts as confounding variable when trying to quantify the performance of such uncertainty measures via correlation. We investigate this for the popular uncertainty measures structure-wise epistemic uncertainty, mean pairwise Dice and volume variation coefficient based on test-time-augmentation, Monte Carlo Dropout and model ensembles. We propose to use a partial correlation coefficient to address structure size as confounding variable and arrive at lower correlation estimates which better reflect the true relationship between segmentation quality and structure-wise uncertainty.

Submission Number: 246