Project

Discipline

Machine learning; Magnetic resonance imaging; Stroke; Mechanical Thrombectomy

Lead
Prof. Dr. Jan GrallaDepartment of Diagnostic and Interventional Neuroradiology, Inselspital, University of Bern
Lay summary
Das STRAY-CATS Projekt Der akute ischämische Schlaganfall ist eine der häufigsten zu Behinderung führenden Erkrankungen in der westlichen Welt. Neue Therapieverfahren, insbesondere die sofortige Eröffnung von akuten Gefässverschlüssen, ermöglichen die erfolgreiche Behandlung von betroffenen Patienten. Während die Verfahren innerhalb von 6 Stunden nach Beginn der Symptome zu Behandlung zugelassen sind, ist es Gegenstand der aktuellen Forschung, Parameter, welche einen Behandlungserfolg auch ausserhalb dieser Zeitfenster zulassen, zu etablieren. Im laufenden Projekt werde multiple, aus akuten MR-Untersuchungen extrahierte Informationen über die Zusammensetzung der Schlaganfälle mittels Methoden des maschinellen Lernens analysiert und Computerprogramme entwickelt, anhand derer die Vorhersage des Risikogewebes (der sogenannten Penumbra) und des bereits irreversibel geschädigten Gewebes ermöglicht wird. Ziel dieses Projektes ist es, dem therapeutisch handelnden Neurologen und Neuroradiologen zusätzliche Informationen über den Kollateralfluss, das rettbare Gewebe und die Aussicht auf den Erfolg einer Thrombektomie zu Verfügung zu stellen. Hierzu werden Datensätze aus mehreren Zentren (Bern, UCLA) und Studien (STAR, Swift Prime) analysiert und anhand eines prospektiven Studienarmes validiert. Abschliessendes Ziel der Studie ist die Etablierung eines neuen bildgebenden und voll automatisierten  Analysetools zur Risikostratifizierung im akuten Schlaganfall.

Direct link to Lay Summary

Last update: 09.11.2017

Natural persons

Number	Title	Start	Funding scheme

Stroke is the second most frequent cause of death and a major cause of disability in industrial countries: in patients who survive, stroke is frequently associated with high socioeconomic costs due to persistent disability. In clinical practice, advanced neuroimaging techniques are increasingly employed for a quick, reliable diagnosis and stratification for therapy. Tissue-at-risk estimation is frequently performed by MRI, with the infarct core being identified as an area of restricted diffusion on diffusion-weighted magnetic resonance imaging (DWI-MRI). The surrounding severly hypoperfused and potentially salvageable tissue tissue (i.e. the “penumbra”) is characterized by its delay in arterial transit time using perfusion-weighted MRI. The clinical image interpretation is routinely performed as a visual analysis done by neuroradiologists and/or neurological stroke experts.There is class I evidence that intravenous thrombolysis is a safe and effective therapy within an estimated time frame of 4.5 h after stroke onset. Very recently, four prospective studies demonstrated the superiority of mechanical thrombectomy in proximal vessel occlusions within a time frame of 6 h after stroke onset. Mechanical thrombectomy has thus become the treatment option of choice to achieve an early and sustained revascularization of proximally occluded vessels in specialized stroke centres2. The recent advent of mechanical thrombectomy has now raised an urgent question that needs to be answered: “can we predict advantageous tissue survival if mechanical thrombectomy is successfully applied compared to the natural course of disease in the presence of sufficient vs. insufficient collaterals?” The availability of a safe, reproducible and reliable information about the expected tissue salvage would allow not only to select patients that would benefit from mechanical thrombectomy, it would further allow to select patients for revascularization in a time window that exceeds 6 h if sufficient collateral flow enables sustained tissue survival. It is essential that indicators for further success of endovascular therapy can be calculated as soon after admission as possible, in order to save as much of the brain tissue as possible: "*time is brain*". Computer-assisted and automated tissue segregation of the infarct core and salvageable penumbra using compound information from multimodal MRI offers a novel and robust standardized solution to this problem: while simple thresholding based on perfusion and diffusion imaging provide only a crude estimate of the tissue at risk, machine-learning approaches based on multimodal MR data overcome the limited accuracy of linear analyses. The proposed machine-learning approach incorporates thus two separate goals, i) to quantify penumbral collateral flow in the acute emergency setting and ii) to identify fingerprints that disentangle salvageable vs. non-salvageable tissue based on machine learning in a “big data” approach based on multiparametric imaging. We will provide means for i) by transforming the interpretation of image features into an interpretation of the underlying stationary flow field. This will allow us to combine information of all available MR imaging data, to quantify the collateral blood flow in the individual patient before and during intervention, and to compare 4D flow patterns at a population level. We will ii) build on our existing predictive models of stroke outcome, incorporating FLAIR and SWI maps and making use of the "big data" that have been acquired during the last years in more than 1000 patients that underwent intraarterial thrombolysis or thrombectomy. Our overall goal is to investigate if, given sufficient training data, predictive maps of the infarction can improve on the current "penumbra" concept as a tool for identifying patients who will have a favourable response to reperfusion therapy.