Fast and uncertainty-aware cerebral cortex morphometry estimation using random forest regression

Abstract

The cortical thickness and curvature of the human brain have proven to be valuable markers to detect and monitor neurodegenerative diseases [1]. Since the computational burden of currently available tools for brain morphometry is very high, this analysis often is only used for retrospective studies and not routinely in the clinics. A first attempt at a clinical use of cortical morphology is reported in [2]. We present an experiment for fast morphometry estimations using Random Forest (RF) regression [3] directly from MR imaging data. An uncertainty-aware voxel-wise, parcellation-wise, and multioutput model was built to estimate the thickness and mean curvature of the human cerebral cortex in 15 minutes instead of many hours for mesh-based tools. Preliminary results on a healthy controls database with 315 subjects show a substantial bias for the voxel-wise prediction, but high scan-rescan robustness, the proposed multi-output-parcellation prediction demonstrates the feasibility of the approach.