Automated Brain Segmentation on Computed Tomographic Images Using Perceptual Loss Based Convolutional Neural Networks

Abstract

Purpose This study aimed to develop a new convolutional neural network-based deep learning (DL) technique for automated brain tissue segmentation from computed tomographic (CT) scans and to evaluate its performance in comparison to magnetic resonance imaging (MRI)-derived segmentations.

Materials and Methods This multicenter retrospective study collected paired CT and MRI data from 199 healthy individuals across two institutions. The data were divided into a training set (n = 100) and an internal test set (n = 50) from one institution, with additional datasets (n = 49) from the second institution for external validation. Ground truth masks for gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) were generated from T1-weighted MR images. A U-Net-based DL model was trained for each of the three brain regions, with a perceptual loss computed from VGG19. Model performance was evaluated by calculating continuous Dice coefficient (cDice), intersection-over-union (IOU), and 95th percentile Hausdorff distance (HD95). Volumetric estimates from CT-based segmentations were compared with MRI-derived volumes using the coefficient of determination (R2), intraclass correlation coefficients (ICC), and Bland-Altman analysis.

Results The DL network trained with the perceptual loss showed superior performance, compared with that trained without the perceptual loss. In internal tests, evaluation scores (without perceptual loss vs. with perceptual loss) were: cDice = 0.717 vs. 0.765 and HD95 = 6.641 mm vs. 6.314 mm in GM; cDice = 0.730 vs. 0.767 and HD95 = 5.841 mm vs. 5.644 mm in WM; and cDice = 0.600 vs. 0.630 and HD95 = 5.641 mm vs. 5.362 mm in CSF, respectively. Volumetric analyses revealed strong agreement between MRI-derived ground truth and CT-based segmentations with R2 = 0.83/0.90 and 0.85/0.87, and ICC = 0.91/0.94 and 0.92/0.93 for GM and WM, respectively, in internal/external tests.

Conclusion The proposed DL method, enhanced with perceptual loss, improves brain tissue segmentation from CT images. This approach shows promise as an alternative to MRIbased segmentation.