This exploratory study demonstrates the feasibility of combining automated lesion segmentation with anatomically informed biomarkers using routine clinical MRI, supporting interpretable stroke outcome modelling and motivating future large-scale validation.
Key Findings
Results
An ensemble of deep learning models (SEALS, NVAUTO, FACTORIZER) achieved a Dice score of 0.82 on the ISLES 2022 test set for stroke lesion segmentation.
The ensemble combined three models: SEALS, NVAUTO, and FACTORIZER.
Evaluated on the ISLES 2022 dataset with 250 training cases and 150 test cases.
Dice score of 0.82 was achieved on ISLES 2022.
Results
Zero-shot external evaluation of the SEALS model alone on ISLES 2024 achieved a Dice score of 0.57.
External evaluation was performed on 149 cases from ISLES 2024.
Only standard MRI sequences were used for this evaluation.
SEALS alone (without the ensemble) was used for external testing.
The zero-shot transfer resulted in a notable performance drop from 0.82 to 0.57 Dice score.
Results
CatBoost achieved the highest point estimates for binary mRS prediction at discharge among the machine learning models evaluated.
CatBoost achieved accuracy of 0.88, F1-score of 0.87, and ROC-AUC of 0.83.
These results are described as 'exploratory analysis' and 'highest point estimates.'
Prediction target was binary modified Rankin Scale (mRS) at discharge.
Results
Key imaging predictors for functional outcome included lesion-CST overlap, lesion volume, surface area, dissimilarity, and contrast.
Lesion-CST overlap was identified as a key predictor, requiring TractSeg-based CST segmentation on single-shell diffusion-weighted imaging.
ADC-based texture features (dissimilarity and contrast) were among the important predictors.
CST segmentation was performed using TractSeg applied to single-shell diffusion-weighted imaging, rather than requiring complex multi-shell diffusion MRI.
TractSeg was used to segment the corticospinal tract.
The approach used single-shell DWI, which is part of routine clinical imaging.
This was designed to improve clinical feasibility compared to methods requiring advanced multi-shell diffusion MRI.
CST integrity and lesion-CST overlap were derived from this segmentation.
Background
Many existing approaches for functional outcome prediction depend on advanced imaging or complex CST segmentation from multi-shell diffusion MRI, limiting clinical feasibility.
The study was motivated by the need for a clinically feasible pipeline based on routine imaging.
Automated lesion segmentation is challenging due to lesion heterogeneity and MRI variability.
Predicting functional outcome at discharge, such as mRS, is important for guiding treatment and rehabilitation.