LST-AI: a Deep Learning Ensemble for Accurate MS Lesion Segmentation.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Automated segmentation of white matter (WM) lesions is critical for multiple sclerosis (MS) assessment and research.
• Existing AI-driven segmentation tools are often proprietary, hindering widespread adoption.
• The LST tool, developed over a decade ago, requires an advanced, open-source successor.

Approach:

• LST-AI, an extension of the LST tool, utilizes an ensemble of three 3D-UNets for enhanced deep learning-based lesion segmentation.
• A composite loss function, combining binary cross-entropy and Tversky loss, addresses the class imbalance between WM lesions and non-lesioned WM.
• Training involved 491 MS patient MRI datasets (T1w and FLAIR) with expert-annotated lesion maps.

Key Points:

• LST-AI achieved superior Dice and F1 scores (>0.62) compared to LST, SAMSEG, and nnUNet (<0.56) on public test data.
• The model demonstrated exceptional performance on the MSSEG-1 challenge, outperforming all competitors with a Dice score of 0.65.
• Detection rates exceeded 75% for lesions between 10mm³ and 100mm³, with improved detection for larger lesion volumes.

Conclusions:

• LST-AI offers a significant advancement in automated WM lesion segmentation for MS research and clinical practice.
• Its open-source nature and versatile deployment options (command-line, Docker, Python script) promote broad accessibility and adoption.
• Researchers using LST are encouraged to transition to LST-AI for improved segmentation performance and lesion characterization.