Detection of diffusely abnormal white matter in multiple sclerosis on multiparametric brain MRI using semi-supervised...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Diffusely abnormal white matter (DAWM) on brain MRI in multiple sclerosis (MS) patients is understudied due to a lack of automated assessment.
• Supervised deep learning (DL) methods require extensive labeled data, posing a challenge for DAWM analysis.

Purpose of the Study:

• To develop and validate a DL-based network (DAWM-Net) for automated segmentation of DAWM, focal lesions, and normal-appearing brain tissues.
• To assess the performance of DAWM-Net compared to a traditional intensity thresholding method.

Main Methods:

• A DL network (DAWM-Net) was trained using semi-supervised learning on limited labeled multiparametric MRI data.
• DAWM-Net segmented DAWM, focal lesions, and normal-appearing brain tissues.
• Segmentation performance was evaluated against expert consensus using Dice Similarity Coefficient (DSC) and volume correlation on an independent test set (N=25).

Main Results:

• DAWM-Net achieved high DSC (>0.93) for normal-appearing brain tissues and (>0.81) for focal lesions.
• DAWM-Net demonstrated improved DAWM segmentation (DSC=0.49 ± 0.12) with moderate volume correlation (ρ=0.52, p<0.01) compared to the previous method (DSC=0.26 ± 0.08, no significant correlation).

Conclusions:

• Deep learning with semi-supervised learning is feasible for automated DAWM segmentation in MS.
• DAWM-Net shows promise as a tool to facilitate future research into the role of DAWM in multiple sclerosis progression and pathology.