Automated Intracranial Vessel Wall Analysis Pipeline for Multi-contrast Multi-platform Applications

Intracranial atherosclerotic disease (ICAD) can lead to ischemic stroke and there is increasing evidence that ICAD, even in the absence of stenosis, is associated with embolic stroke of undetermined source (ESUS). Vessel wall magnetic resonance imaging (MRI) of the intracranial vasculature is increasingly in demand to diagnose such ESUS patients so that appropriate treatment can be administered. Multi-contrast intracranial vessel wall (IVW) MRI with efficient analysis is currently recommended by the American Society of Neuroradiology to diagnose various vessel wall pathologies including ICAD. While this urgent need for IVW has stimulated availability of 3D spin echo sequences on major scanner platforms (VISTA on Philips, SPACE on Siemens and CUBE on GE), there are no efficient and effective methods for clinicians to analyze the multi-contrast IVW MRI sequences that can be obtained on modern clinical MRI scanners. Quantitative measurements of the vessel wall across scanner platforms are also required to enable multi-center studies for ICAD assessment in ESUS. Variability in sequence implementation affects multi-center studies on multiple scanner platforms and must be overcome to enable robust IVW measurements. Therefore, we propose to develop an automated IVW analysis pipeline for multi-contrast multi-platform application using a domain adaptive and deep learning approach. We have pioneered multiple semiautomatic approaches (3D-registration, artery tracing, artery labeling, multi-planar reformatting, vessel wall segmentation, multi-contrast feature identification) towards vessel wall quantification. Leveraging this expertise, we will develop a novel artificial intelligence (AI) empowered multiplanar viewing for artery characterization (AI-MOCHA) pipeline as follows: In Aim 1 we will construct the MOCHA pipeline and train AI-MOCHA using transfer learning from labeled IVW images and test AI-MOCHA against radiologist labeled IVW from ICAD patients. We will also test whether AI-MOCHA improves the inter-scan and inter-reader reproducibility of IVW image analysis. In Aim 2 we will develop domain adaptation to overcome scanner-platform differences in IVW images and develop a Domain Adaptive AI-MOCHA. We will then show that domain adaptive AI-MOCHA improves the inter-scan and inter-reader reproducibility of IVW image analysis over AI-MOCHA. In Aim 3, we will test the hypothesis that non-stenotic ICAD is more frequently detected by AI-MOCHA in the vascular territory of ESUS than in other territories using Domain Adaptive AI-MOCHA. To achieve this, we will scan 65 ESUS subjects each on the Philips, Siemens and GE 3T scanner platforms in a multi-center setting (three different hospitals) and demonstrate the utility of domain adaptive AI-MOCHA for robust and efficient IVW analysis. In doing so, we will not only establish the importance of non-stenotic ICAD in ESUS but also develop a clinically applicable IVW analysis pipeline for monitoring ICAD progression that will assist in optimizing medical therapies in individual patients. Further, the pipeline will provide a rapid, reliable tool for identifying patients who do not respond to conventional therapies for clinical trial participation.