Multiple sclerosis cortical lesion detection with deep learning at ultra-high-field MRI

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Manual segmentation of multiple sclerosis cortical lesions (CLs) is time-consuming and suffers from moderate inter-rater reliability.
• Accurate CL detection is crucial for MS diagnosis and monitoring.

Purpose of the Study:

• To develop and evaluate a deep-learning framework (CLAIMS) for automated detection and classification of MS CLs using 7T MRI.
• To compare CLAIMS performance against state-of-the-art methods and manual segmentation.

Main Methods:

• Developed CLAIMS, a deep learning framework for CL detection and classification.
• Evaluated CLAIMS on two independent 7T MRI datasets with varying resolutions and contrasts.
• Compared CLAIMS with MSLAST, a state-of-the-art technique, using lesion-wise detection rates and correlation with disability measures.

Main Results:

• CLAIMS achieved a CL true positive rate of 74% with a 30% false positive rate, comparable to multi-contrast models.
• Detection rates were 83% for leukocortical, 70% for subpial, and 53% for intracortical lesions.
• CLAIMS outperformed MSLAST on an external dataset, showing a 71% lesion-wise detection rate versus 48% for MSLAST.

Conclusions:

• The CLAIMS framework provides accurate and automated detection of MS CLs using 7T MRI.
• CLAIMS demonstrates superior generalizability and performance across different scanners and protocols compared to existing methods.
• CLAIMS has the potential to support clinical decisions in the diagnosis and differential diagnosis of MS.