Abstract
BACKGROUND AND PURPOSE: Despite high interest in machine-learning algorithms for automated segmentation of MRIs of patients with brain tumors, there are few reports on the variability of segmentation results. The purpose of this study was to obtain benchmark measures of repeatability for a widely accessible software program, BraTumIA (Versions 1.2 and 2.0), which uses a machine-learning algorithm to segment tumor features on contrast-enhanced brain MR imaging. MATERIALS AND METHODS: Automatic segmentation of enhancing tumor, tumor edema, nonenhancing tumor, and necrosis was performed on repeat MR imaging scans obtained approximately 2 days apart in 20 patients with recurrent glioblastoma. Measures of repeatability and spatial overlap, including repeatability and Dice coefficients, are reported. RESULTS: Larger volumes of enhancing tumor were obtained on later compared with earlier scans (mean, 26.3 versus 24.2 mL for BraTumIA 1.2; P < .05; and 24.9 versus 22.9 mL for BraTumIA 2.0, P < .01). In terms of percentage change, repeatability coefficients ranged from 31% to 46% for enhancing tumor and edema components and from 87% to 116% for nonenhancing tumor and necrosis. Dice coefficients were highest (>0.7) for enhancing tumor and edema components, intermediate for necrosis, and lowest for nonenhancing tumor and did not differ between software versions. Enhancing tumor and tumor edema were smaller, and necrotic tumor larger using BraTumIA 2.0 rather than 1.2. CONCLUSIONS: Repeatability and overlap metrics varied by segmentation type, with better performance for segmentations of enhancing tumor and tumor edema compared with other components. Incomplete washout of gadolinium contrast agents could account for increasing enhancing tumor volumes on later scans.