Optimizing Choroid Plexus Segmentation using nnU-Net: Comparisons among Multimodal MRI Input Combinations

Motivation: Accurate segmentation of the choroid plexus (CP) is gaining importance for monitoring neurodegenerative/neuroinflammatory diseases as a reliable imaging biomarker, yet existing automated segmentation methods show limited accuracy. Goal(s): To develop a high-precision CP segmentation method by using a combination of multimodal MRI sequences. Approach: We developed an nnU-Net-based deep learning model trained on 119 expert-annotated brain MRI datasets, evaluating four different input combinations of 3D-T1WI, 3D-FLAIR, and 3D-CE-T1WI. Results: Integration of all three MRI sequences achieved the best performance (Dice: 0.787±0.059, IoU: 0.652±0.078) with excellent volumetric reliability (r=0.861, p<0.001). Impact: Our findings demonstrate that 3D-CE-T1WI significantly improves the accuracy of CP segmentation. This multimodal approach would facilitate more reliable volumetric analysis in both research and clinical settings.