A Multiple Sclerosis MRI Dataset with Tri-Mask Annotations for Lesion Segmentation.

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Accurate multiple sclerosis (MS) lesion segmentation in MRI is crucial for clinical management.
• Existing datasets lack geographic diversity and struggle to differentiate normal from pathological white matter hyperintensities.

Purpose of the Study:

• To introduce MS3SEG, a novel, publicly available dataset for MS lesion segmentation.
• To address the challenge of distinguishing benign white matter hyperintensities from MS lesions.
• To facilitate research on robust and clinically relevant MS segmentation algorithms.

Main Methods:

• Developed MS3SEG dataset with 100 MS patients from an Iranian cohort using a 1.5T Toshiba scanner.
• Acquired multi-sequence MRI (T1w, T2w, T2-FLAIR) in axial and sagittal planes.
• Implemented a novel tri-mask annotation framework (ventricles, normal, abnormal hyperintensities) with automated quality control on axial T2-FLAIR images.

Main Results:

• Baseline validation with U-Net, U-Net++, UNETR, and Swin UNETR models.
• U-Net achieved Dice coefficients of 0.7469 for binary and 0.6686 for multi-class abnormal hyperintensity segmentation.
• Demonstrated the dataset's quality and utility for training segmentation algorithms.

Conclusions:

• MS3SEG provides a valuable resource for developing anatomically-aware MS segmentation algorithms.
• The dataset aids research into algorithm robustness against real-world acquisition variations.
• Enables clinically relevant distinction between normal and abnormal white matter hyperintensities, enhancing MS diagnosis and monitoring.