ASCHOPLEX: A generalizable approach for the automatic segmentation of choroid plexus

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• The Choroid Plexus (ChP) is crucial for brain homeostasis and cerebrospinal fluid production.
• Accurate segmentation of the ChP is challenging but vital for studying neurological and psychiatric disorders.
• Existing segmentation methods lack accuracy and reproducibility for large-scale studies.

Purpose of the Study:

• To introduce ASCHOPLEX, a novel deep learning tool for automated Choroid Plexus segmentation.
• To enable accurate and reproducible ChP volume estimation from structural MRI data.
• To improve the analysis of ChP in neurological and psychiatric conditions.

Main Methods:

• ASCHOPLEX utilizes 3D UNet, UNETR, and DynUNet architectures.
• The tool was trained on 128 T1-weighted MRI scans (controls and Multiple Sclerosis patients).
• Fine-tuning (ASCHOPLEXtune) was performed on 77 T1-weighted PET/MRI scans (controls and depressed patients).

Main Results:

• ASCHOPLEX achieved superior performance over FreeSurfer and Gaussian Mixture Model.
• Achieved Dice Coefficients of 0.80 (ASCHOPLEX) and 0.78 (ASCHOPLEXtune).
• Demonstrated low ChP volume error: 9.22% (ASCHOPLEX) and 9.23% (ASCHOPLEXtune).

Conclusions:

• ASCHOPLEX provides highly accurate and reliable Choroid Plexus segmentations.
• The tool ensures reproducible results for ChP volume estimation.
• ASCHOPLEX is a valuable tool for neuroimaging research involving the Choroid Plexus.