LST-AI: A deep learning ensemble for accurate MS lesion segmentation.

Area of Science:

• Medical Imaging and Radiology
• Artificial Intelligence in Medicine
• Neuroscience and Neurology

Background:

• Accurate segmentation of white matter (WM) lesions is critical for multiple sclerosis (MS) clinical assessment and research.
• Existing AI-based lesion segmentation tools are often proprietary and difficult to implement.
• The previously developed LST tool requires an advanced, open-source successor.

Purpose of the Study:

• To introduce LST-AI, an open-source, deep learning-based extension of the LST tool for automated brain white matter lesion segmentation in MS.
• To improve upon existing lesion segmentation methods by addressing data imbalance and lesion heterogeneity.
• To provide a user-friendly, adaptable tool for researchers and clinicians.

Main Methods:

• Developed LST-AI using an ensemble of three 3D U-Nets, trained on 491 in-house MS MRI datasets (T1-weighted and FLAIR).
• Employed a composite loss function (binary cross-entropy and Tversky loss) to enhance segmentation of heterogeneous MS lesions.
• Integrated a lesion location annotation tool based on the 2017 McDonald criteria and included subcortical labeling.

Main Results:

• LST-AI achieved superior performance, with Dice and F1 scores exceeding 0.62, outperforming LST, SAMSEG, and nnUNet (all <0.56).
• Demonstrated exceptional performance on the MSSEG-1 challenge dataset (Dice: 0.65, F1: 0.63), surpassing all competing models.
• Achieved a lesion detection rate of >75% for lesions between 10 mm³ and 100 mm³, with detection rates increasing with lesion volume.

Conclusions:

• LST-AI represents a significant advancement in automated white matter lesion segmentation for MS.
• The open-source availability and superior performance make LST-AI a recommended replacement for LST and other existing tools.
• LST-AI facilitates broader adoption and diverse applications across multiple platforms for MS research and clinical practice.