Benchmarking and Explaining Deep Learning Cortical Lesion MRI Segmentation in Multiple Sclerosis.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomarker Discovery

Background:

• Cortical lesions (CLs) are key biomarkers in multiple sclerosis (MS) but are difficult to detect and segment in MRI.
• Current methods lack standardization, hindering routine clinical use.

Purpose of the Study:

• To establish a multi-centric benchmark for automated CL detection and segmentation in MRI.
• To adapt and evaluate the nnU-Net framework for improved CL analysis.

Main Methods:

• Utilized 656 multi-institutional MRI scans (3T and 7T) with MP2RAGE and MPRAGE sequences.
• Employed the self-configuring nnU-Net framework with tailored adaptations for CL detection.
• Performed out-of-distribution testing to assess model generalization.

Main Results:

• Achieved an F1-score of 0.64 for in-domain and 0.5 for out-of-domain CL detection.
• Analyzed model features and errors to understand AI decision-making.
• Identified impacts of data variability and protocol differences on performance.

Conclusions:

• The proposed approach demonstrates robust CL detection capabilities, addressing limitations in current MS diagnostics.
• Findings offer recommendations for overcoming barriers to clinical adoption of automated MRI analysis.
• Publicly accessible code and models will enhance reproducibility and future research.