Localized prediction of tissue outcome in acute ischemic stroke patients using diffusion- and perfusion-weighted MRI datasets

Abstract

Background: An accurate prediction of tissue outcome in acute ischemic stroke patients is of high interest for treatment decision making. To date, various machine learning models have been proposed that combine multi-parametric imaging data for this purpose. However, most of these machine learning models were trained using voxel information extracted from the whole brain, without taking differences in susceptibility to ischemia into account that exist between brain regions. The aim of this study was to develop and evaluate a local tissue outcome prediction approach, which makes predictions using locally trained machine learning models and thus accounts for regional differences.

Material and methods: Multi-parametric MRI data from 99 acute ischemic stroke patients were used for the development and evaluation of the local tissue outcome prediction approach. Diffusion (ADC) and perfusion parameter maps (CBF, CBV, MTT, Tmax) and corresponding follow-up lesion masks for each patient were registered to the MNI brain atlas. Logistic regression (LR) and random forest (RF) models were trained employing a local approach, which makes predictions using models individually trained for each specific voxel position using the corresponding local data. A global approach, which uses a single model trained using all voxels of the brain, was used for comparison. Tissue outcome predictions resulting from the global and local RF and LR models, as well as a combined (hybrid) approach were quantitatively evaluated and compared using the area under the receiver operating characteristic curve (ROC AUC), the Dice coefficient, and the sensitivity and specificity metrics.

Results: Statistical analysis revealed the highest ROC AUC and Dice values for the hybrid approach. With 0.872 (ROC AUC; LR) and 0.353 (Dice; RF), these values were significantly higher (p < 0.01) than the values of the two other approaches. In addition, the local approach achieved the highest sensitivity of 0.448 (LR). Overall, the hybrid approach was only outperformed in sensitivity (LR) by the local approach and in specificity by both other approaches. However, in these cases the effect sizes were comparatively small.

Conclusion: The results of this study suggest that using locally trained machine learning models can lead to better lesion outcome prediction results compared to a single global machine learning model trained using all voxel information independent of the location in the brain.

Fig 1. Final infarct outcome predictions for a selected patient.

Images of the final infarct outcome predictions (first and third row) and binarized masks (second and fourth row) for the global, local, and hybrid (left to right) LR and RF (top to bottom) models for a selected patient and the corresponding true follow-up lesion outcome shown on the far right. In areas where the global models underestimates the lesion, the local models show higher infarct probabilities leading to better fits of the binary prediction masks, with the true follow-up lesion outcome; for both, LR (A) and RF (B). In addition, the global LR and RF models show a smooth coherent infarct prediction, whereas the local approach, especially the RF model, is slightly more scattered. Nevertheless, the dispersion of the local approach is concentrated on the actual infarct regions, so that the hybrid prediction is not only smoother than the local approaches, but also leads to the overall best qualitative results.