Deep Learning-Based Corpus Callosum Segmentation Applied Directly on Diffusion Tensor Imaging Data: a Volumetric and Midsagittal Analysis

The corpus callosum (CC) has been widely studied using various imaging techniques, including diffusion tensor imaging (DTI), for microstructural analysis. These analyses often require segmentation of the CC, which is commonly performed on T1-weighted images, necessitating registration to diffusion space. To overcome this limitation, we propose an automated method for volumetric CC segmentation directly in diffusion space using a U-Net architecture, a convolutional neural network widely applied in medical image segmentation. On a test set of 50 subjects with manual annotations, our approach outperformed the MRICloud and FreeSurfer atlas-based methods, achieving a mean Dice of 0.850 and average Hausdorff distance (AVD) of 0.195. We also obtained midsagittal CC segmentations from the volumetric results. Compared with DTI-based methods in the inCCsight tool, i.e., watershed-based segmentation and reproducible objective quantification scheme, our method showed superior performance on an in-house dataset, with a mean Dice of 0.916 and AVD of 0.087. However, midsagittal results were similar among methods (mean Dice of approximately 0.9) on the Human Connectome Project dataset, indicating the potential for improvement. The proposed method enables detailed volumetric analysis of the anatomical and microstructural properties of the CC while maintaining compatibility with existing two-dimensional studies. Integrated into the inCCsight tool, it can enhance reproducibility and has potential applications in early diagnosis, disease monitoring, and neurosurgical planning.