Accurate brain tissue segmentation is essential for constructing individualized head models in noninvasive brain stimulation. This study investigated how loss function design influences segmentation performance using a large dataset of over 500 T1-weighted MRIs. A 3D U-Net was trained to segment the skin, skull, cerebrospinal fluid (CSF), gray matter (GM), and white matter (WM) using either Dice or combined Dice + Cross-Entropy (Dice+CE) losses. Quantitative results showed that Dice+CE achieved higher overall accuracy (0.8577 vs. 0.8550) and mean IoU (0.2863 vs. 0.2716), with notable improvements in low-contrast tissues such as skin, skull, CSF, and GM. WM slightly favored Dice due to its large and homogeneous structure. Qualitative analysis indicated clearer tissue boundaries and fewer misclassifications with Dice+CE. These findings demonstrate that combining Dice and Cross-Entropy losses enhances segmentation accuracy and provide evidence for optimizing brain tissue segmentation model performance.
Kim et al. (Wed,) studied this question.