Ischemic Stroke Thrombus Characterization through Quantitative Magnetic Resonance Imaging

Spencer D. Christiansen, The University of Western Ontario

Abstract

Stroke is a pervasive, devastating disease and remains one of the most challenging conditions to treat. In particular, risk of recurrence is dramatically heightened after a primary stroke and requires urgent preventative therapy to effectively mitigate. However, the appropriate preventative therapy depends on the underlying source of the stroke, known as etiology, which is challenging to determine promptly. Current diagnostic tests for determining etiology underwhelm in both sensitivity and specificity, and in as much as 35% of cases etiology is never determined. In ischemic stroke, the composition of the occluding thrombus, specifically its red blood cell (RBC) content, has been shown to be indicative of etiology but remains largely ignored within clinical practice. Currently, composition can only be quantified through histological analysis, an involved process that can be completed in only the minority of cases where a thrombus has been physically retrieved from the patient during treatment.

The goal of this thesis is to develop a quantitative MR imaging method which is capable of non-invasive prediction of ischemic stroke etiology through assessment of thrombus RBC content. To achieve this goal, we employed quantitative MR parameters that are sensitive to both RBC content and oxygenation, R₂^* and quantitative susceptibility mapping (QSM), as well as fat fraction (FF) mapping, and evaluated the ability of modern artificial intelligence techniques to form predictions of RBC content and etiology based on these quantitative MR parameters alone and in combination with patient clinical data.

First, we performed an in vitro blood clot imaging experiment, which sought to explicitly define the relationship between clot RBC content, oxygenation and our quantitative MR parameters. We show that both R₂^* and QSM are sensitive to RBC content and oxygenation, as expected, and that the relationship between R₂^* and QSM can be used to predict clot RBC content independent of oxygenation status, a necessary step for stroke thrombus characterization where oxygenation is an unknown quantity.

Second, we trained a deep convolutional neural network to predict histological RBC content from ex vivo MR images of ischemic stroke thrombi. We demonstrate that a convolutional neural network is capable of RBC content prediction with 66% accuracy and 8% mean absolute error when trained on a limited thrombus dataset, and that prediction accuracy can be improved to up to 74% through data augmentation.

Finally, we used a random forest classifier to predict clinical stroke etiology using the same ex vivo thrombus MR image dataset. Here, quantitative thrombus R₂^*, QSM and FF image texture features were used to train the classifier, and we demonstrate that this method is capable of accurate etiology prediction from thrombus texture information alone (accuracy = 67%, area under the curve (AUC) = 0.68), but that when combined with patient clinical data the performance of the classifier improves dramatically to an accuracy and AUC of 93% and 0.89, respectively.

Together, the works presented in this thesis offer extensive ex vivo evidence that quantitative MR imaging is capable of effective stroke thrombus etiology characterization. Such a technique could enable clinical consideration of thrombus composition and bolster stroke etiology determination, thereby improving the management and care of acute ischemic stroke patients.