Deep Learning Informed by Contrast-enhanced MRI Allows for Precise Segmentation of the Choroid Plexus in Non-contrast Data

RATIONALE AND OBJECTIVES: The choroid plexus (ChP) is involved in cerebrospinal fluid (CSF) production and neuroimmunology. Its precise delineation in neuroimaging is challenging. We developed a deep learning-based approach for automated segmentation of the ChP in contrast-enhanced(CE) brain magnetic resonance imaging (MRI), applied it to CE and non-contrast(NC) MRI, and compared the performance to an established segmentation algorithm. MATERIALS AND METHODS: We relied on cerebral NC and CE 3D datasets of 199 patients prior to stereotactic intervention (mean age: 57 years, age range: 1-83 years, 43% women). The lateral ventricle ChP was manually segmented on 3D T1w post-contrast images to serve as ground truth, and a deep neural patchwork (DNP) was trained based on 149 randomly chosen datasets using information of pre- and post-contrast T1w and T2w images. The DNP was applied to the remaining 50 datasets for each contrast independently and compared with FreeSurfer-based ChP segmentation based on NC T1w. RESULTS: The DNP algorithm successfully segmented the ChP in the training datasets and reached a mean dice coefficient of 0.81 ± 0.05 in T1w post-contrast. Upon application on NC T1w or T2w images, sufficient quality with mean dice coefficients of 0.71 ± 0.05 and 0.70 ± 0.05 was noted. DNP segmentations on any contrast significantly outperformed FreeSurfer (0.38 ± 0.05; all p < .001). CONCLUSION: The deep learning-based segmentation algorithm delineating the ChP in CE and NC MRI datasets, might allow for a more reliable insight into the long-term evolution of physiological and pathological processes.