Related Experiment Video
Updated: Sep 2, 2026
Manual Segmentation of the Human Choroid Plexus Using Brain MRI
Published on: December 15, 2023
Ultrafast (0.5-Minute) Deep Learning Based Choroid Plexus Segmentation: Links to Cognitive Impairment and Glymphatic
Hyunsoh Park1, Woojin Jung1, Sung-Hye You2
1From the Korea University College of Medicine (H.P., H.L.), Republic of Korea; AIRS Medical (W.J.), Republic of Korea; Department of Radiology (S.-H.Y., B.K., M.H., B.K. K., J.H.S.), Anam Hospital, Radiology (S-I.S., I.R., H.N.J.), Guro Hospital, Radiology (Y.L.), Ansan Hospital, Korea University College of Medicine, Republic of Korea.
Background And Purpose:
Choroid plexus volume (CPV) may reflect cognitive impairment and glymphatic dysfunction. However, its clinical use is limited by time-consuming segmentation. We developed an ultrafast deep learning method for automated CPV quantification using 3D T1-weighted images (T1WI).
Materials And Methods:
This retrospective study included 697 patients who underwent brain MRI for cognitive impairment. The model was developed using 3D T1WI from 103 patients as an ensemble of three 2.5D U-Nets in the axial, coronal, and sagittal planes. Imaging validation was performed for 20 patients by comparing manual segmentation, FreeSurfer, and the proposed method. Clinical validation included 574 patients classified as subjective cognitive impairment (SCI), early mild cognitive impairment (MCI), late MCI, or Alzheimer disease (AD). ComBat harmonization was applied to reduce scanner-related effects. Accuracy, processing time, group differences, and correlations with the diffusion tensor imaging analysis along the perivascular space (DTI-ALPS) index were assessed.
Results:
After ComBat harmonization, the deep learning-based model was markedly faster than FreeSurfer (0.4 vs. 188.0 minutes; P<0.001), and its CPV measurements showed no significant difference from manual segmentation (1.89 vs. 1.84 mL; P=1.00). CPV/intracranial volume (ICV)×10³ increased progressively across cognitive stages, from 1.02 in SCI to 1.16 in AD (P=0.004). CPV/ICV measured by the deep learning-based model was negatively correlated with the DTI-ALPS index (ρ=-0.256, P<0.001).
Conclusion:
An ultrafast deep learning method enables automated CPV quantification and provides clinically relevant CPV/ICV measurements for neurodegenerative disease assessment.
