Leveraging data-driven segmentation uncertainty estimates as potential diagnostic indicators.

BACKGROUND

The ability to quantify uncertainty is a desirable feature of explainable artificial intelligence systems. In the case of segmentation models, however, practical clinical applications for uncertainty estimates have been under-explored and most studies are limited to simple single-class problems. The goal of this work is to expand uncertainty analysis to more complex segmentation problems and to evaluate the feasibility of using uncertainty information as a marker of potential pathology.

METHODS

Using the publicly available TotalSegmentator model, and a custom Multi-Organ Test-Time Augmentation (TTA) pipeline, abdominal segmentation uncertainty was characterized across 14 total organs and 8 total augmentations. Uncertainty was defined as the average entropy across the predicted probabilities for the augmented versions of the provided input image and quantified at an organ and augmentation level by averaging the entropy values within a dilated mask of the target structure. Uncertainty performance was characterized in 872 total abdominal computed tomography cases, and an additional 489 cases were used to evaluate the sensitivity of uncertainty estimates to pathology in the pancreas, liver, colon and kidneys where ground truth was available.

RESULTS

Uncertainty estimates were found to be significantly higher in pathologic tissue versus healthy tissue regardless of the organ considered (pancreas: P=1.79e-05; liver: P=8.83e-23; colon: P=1.36e-26; right kidney: P=2.49e-57; left kidney: P=1.43e-50). Receiver operating characteristic analysis of the uncertainty maps highlighted fair voxel-level pathology detection performance in the pancreas [area under the curve (AUC): 0.7097], colon (AUC: 0.7316), and left kidney (AUC: 0.7605), good performance in the right kidney (AUC: 0.8134) and excellent performance in the liver (AUC: 0.8885).

CONCLUSIONS

TTA-derived uncertainty estimates have been proved useful as a tool for detecting pathology across multiple organ systems. With future development, a tool such as this could prove useful in early screening for pathology, or in an active learning scheme to identify regions in need of manual intervention.