Evaluating Intra- and Inter- scan session consistency of Machine Learning Multi Organ Segmentation

Abstract

Motivation: Machine learning segmentation is becoming more widely used and applied in longitudinal studies to assess changes in organ volume, with organ masks applied to evaluate quantitative metrics. Goal(s): To evaluate the abdominal organ segmentation consistency of nnU-Net. Approach: nnU-Net segmentation models were trained to segment abdominal organs. These models were applied to assess intra-session and inter-session repeatability of organ volume measures. Results: Intra-scan session variance for all organs was extremely low (<1.4 %). Inter-scan variance was low across most organs (<6 %), with the exceptions of some adipose tissue segmentations and kidney cortex/medulla segmentations. Impact: Assessment of intra- and inter- scan session reproducibility provides an understanding of the detectable change in organ volume in longitudinal studies and the accuracy of masks used to derive associated quantitative metrics.

Ähnliche Arbeiten