Automated claustrum segmentation in human brain MRI using deep learning.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• The claustrum plays a key role in mammalian brain function, but in vivo human studies are limited.
• Its delicate, sheet-like structure between the insular cortex and putamen challenges conventional segmentation.
• Deep learning (DL) offers automated segmentation for complex subcortical brain structures.

Purpose of the Study:

• To develop and validate a multi-view Deep Learning (DL) approach for automated segmentation of the human claustrum.
• To assess the performance and generalizability of the DL model for claustrum segmentation.

Main Methods:

• A multi-view DL model was developed to segment the claustrum in T1-weighted MRI scans.
• The method was trained and evaluated on 181 individuals, with manual annotations by an expert neuroradiologist serving as the reference standard.
• Cross-validation and leave-one-scanner-out evaluations were performed to assess performance and transferability.

Main Results:

• The DL model achieved high segmentation accuracy, with median volumetric similarity of 93.3%, Hausdorff distance of 1.41 mm, and Dice score of 71.8%.
• Performance was comparable or superior to human intra-rater reliability.
• The model demonstrated good transferability to unseen scanners, though with slightly reduced performance.
• Multi-view information improved DL-based claustrum segmentation, with an optimal training set size of approximately 75 MRI scans.

Conclusions:

• The developed algorithm enables robust and automated segmentation of the human claustrum from MRI scans.
• This method has significant potential to facilitate future research on the human claustrum.
• The software and models are publicly available to support the research community.