Automated Multi-Modal MRI Segmentation of Stroke Lesions and Corticospinal Tract Integrity for Functional Outcome...

Area of Science:

• Neurology
• Medical Imaging
• Machine Learning

Background:

• Stroke significantly contributes to long-term disability, necessitating accurate prediction of functional outcomes like the modified Rankin Scale (mRS).
• Current prediction methods often rely on advanced neuroimaging or complex corticospinal tract (CST) segmentation, limiting clinical applicability.
• Automated lesion segmentation faces challenges due to lesion variability and MRI inconsistencies.

Purpose of the Study:

• To develop and validate a clinically feasible multimodal MRI pipeline for predicting stroke functional outcomes at discharge.
• To assess the utility of routine MRI sequences for lesion segmentation and corticospinal tract (CST) analysis.
• To identify key imaging biomarkers for stroke outcome prediction.

Main Methods:

• Utilized deep learning models (SEALS, NVAUTO, FACTORIZER) for lesion segmentation on the ISLES 2022 dataset and external validation on ISLES 2024.
• Performed CST segmentation using TractSeg on single-shell diffusion-weighted imaging.
• Extracted imaging biomarkers including lesion volume, shape, texture, CST integrity, and lesion-CST overlap.
• Trained machine learning models (e.g., CatBoost) for binary mRS prediction using these biomarkers.

Main Results:

• The deep learning ensemble achieved a Dice score of 0.82 on the ISLES 2022 dataset for lesion segmentation.
• External validation on ISLES 2024 yielded a Dice score of 0.57.
• The CatBoost model demonstrated strong predictive performance with accuracy 0.88, F1-score 0.87, and ROC-AUC 0.83.
• Significant predictors included lesion-CST overlap, lesion volume, surface area, dissimilarity, and contrast.

Conclusions:

• This study demonstrates the feasibility of a multimodal MRI pipeline using routine imaging for stroke outcome prediction.
• The approach combines automated lesion segmentation with anatomically informed biomarkers for interpretable stroke modeling.
• Findings support the potential for large-scale validation and clinical implementation of this pipeline.