Reproducibility of radiomic features of the brain on ultrahigh-resolution MRI at 7 Tesla: a comparison of different segmentation techniques

ABSTRACT Objectives To determine the impact of segmentation techniques on radiomic features extracted from ultrahigh-field (UHF) MRI of the brain. Materials and Methods Twenty-one 7T MRI scans of the brain, including a 3D magnetization-prepared two rapid acquisition gradient echo (MP2RAGE) T1-weighted sequence with an isotropic 0.63 mm³ voxel size, were analyzed. Radiomic features (histogram, texture, and shape; total n=101) from six brain regions -cerebral gray and white matter, basal ganglia, ventricles, cerebellum, and brainstem-were extracted from segmentation masks constructed with four different techniques: the iGT (reference standard), based on a custom pipeline that combined automatic segmentation tools and expert reader correction; the deep-learning algorithm Cerebrum-7T; the Freesurfer-v7 software suite; and the Nighres algorithm. Principal components (PCs) were calculated for histogram and texture features. To test the reproducibility of radiomic features, intraclass correlation coefficients (ICC) were used to compare Cerebrum-7, Freesurfer-v7, and Nighres to the iGT, respectively. Results For histogram PCs, median ICCs for Cerebrum-7T, Freesurfer-v7, and Nighres were 0.99, 0.42, and 0.11 for the gray matter; 0.84, 0.25, and 0.43 for the basal ganglia; 0.89, 0.063, and 0.036 for the white matter; 0.84, 0.21, and 0.33 for the ventricles; 0.94, 0.64, and 0.93 for the cerebellum; and 0.78, 0.21, and 0.53 for the brainstem. For texture PCs, median ICCs for Cerebrum-7T, Freesurfer-v7, and Nighres were 0.95, 0.21, and 0.15 for the gray matter; 0.70, 0.36, and 0.023 for the basal ganglia; 0.91, 0.25, and 0.023 for the white matter; 0.80, 0.75, and 0.59 for the ventricles; 0.95, 0.43, and 0.86 for the cerebellum; and 0.72, 0.39, and 0.46 for the brainstem. For shape features, median ICCs for Cerebrum-7T, FreeSurfer-v7, and Nighres were 0.99, 0.91, and 0.36 for the gray matter; 0.89, 0.90, and 0.13 for the basal ganglia; 0.98, 0.91, and 0.027 for the white matter; 0.91, 0.91, and 0.36 for the ventricles; 0.80, 0.68, and 0.47 for the cerebellum; and 0.79, 0.17, and 0.15 for the brainstem. Conclusions Radiomic features in UHF MRI of the brain show substantial variability depending on the segmentation algorithm. The deep learning algorithm Cerebrum-7T enabled the highest reproducibility. Dedicated software tools for UHF MRI may be needed to achieve more stable results.