Objectives: To develop an algorithm to segment and obtain an estimate of total intracranial volume (tICV) from computed tomography (CT) images.
Materials and methods: Thirty-six CT examinations from 18 patients were included. Ten patients were examined twice the same day and eight patients twice six months apart (these patients also underwent MRI). The algorithm combines morphological operations, intensity thresholding and mixture modelling. The method was validated against manual delineation and its robustness assessed from repeated imaging examinations. Using automated MRI software, the comparability with MRI was investigated. Volumes were compared based on average relative volume differences and their magnitudes; agreement was shown by a Bland-Altman analysis graph.
Results: We observed good agreement between our algorithm and manual delineation of a trained radiologist: the Pearson's correlation coefficient was r = 0.94, tICVml[manual] = 1.05 × tICVml[automated] - 33.78 (R(2) = 0.88). Bland-Altman analysis showed a bias of 31 mL and a standard deviation of 30 mL over a range of 1265 to 1526 mL.
Conclusions: tICV measurements derived from CT using our proposed algorithm have shown to be reliable and consistent compared to manual delineation. However, it appears difficult to directly compare tICV measures between CT and MRI.
Key points: • Automated estimation of tICV is in good agreement with manual tracing. • Consistent tICV estimations from repeated measurements demonstrate the robustness of the algorithm. • Automatically segmented volumes seem less variable than those from manual tracing. • Unbiased and automated tlCV estimation is possible from CT.
Keywords: Computed tomography; Magnetic resonance imaging; Maximum likelihood estimator; Skull stripping; Total intracranial volume.