Fast MR Imaging And Intelligent Analysis For Early Detection Of Plaques For Stroke

Stroke is a leading cause of death in China,and 79% of strokes are ischemic stroke,mainly due to rupture of atherosclerotic plaques in the intracranial and carotid arteries.Early detection and assessment of vulnerable plaques may help elucidate stroke etiology and allow for therapeutic intervention before severe cerebrovascular events occur.Magnetic resonance imaging（MRI）has the unique advantages of high soft-tissue contrast and no ionizing radiation,and its diagnostic value for the arterial plaque has been basically recognized clinically.With the popularization of high-end magnetic resonance equipment,the MRI for stroke high-risk populations carrying out comprehensive plaque screening and etiological exploration may become important methods for the prevention and therapeutic of stroke in the future.However,the popularization applications of MRI plaque imaging still face two major bottlenecks:（1）the acquisition time is too long.It exceeds the limit that most examiners tolerate and easily produces the motion artifacts;（2）professional doctors take a long time to read and diagnostic due to the huge amount of data of the three-dimensional（3D）highresolution magnetic resonance vessel wall imaging,which hinders the progress of clinical popularization.For the above-mentioned problems,this dissertation aims to study the fast MRI and the intelligent analysis of plaques,to provide a new tool for the assessment and prediction of stroke risk.To achieve the fast and high-resolution imaging of vessel walls,we developed a novel 3D imaging sequence of the vessel walls and a fast reconstruction algorithm based on machine learning.We further explored artificial intelligence diagnosis-related vessel centerline tracking algorithms,deep learning segmentation networks,etc.,to achieve the whole intelligent process from plaque imaging to diagnosis.The main results and structure of this dissertation are as following.Firstly,we developed a new imaging sequence of the blood vessel walls due to high-quality vessel wall images being prerequisites for intelligent diagnosis.The dissertation employed an iterative optimization method to calculate RF pulse flip.It can optimize variable flip mode according to the signal characteristics of the variety and complex tissues.Based on this algorithm,we designed a new 3D modulated flip angle technique in refocused imaging with an extended echo train（MATRIX）imaging sequence.Furthermore,the DANTE（Delays Alternating with Nutation for Tailored Excitation）technology was utilized as a backup pulse for the suppression of blood flow and cerebrospinal fluid.We investigated the motion-sensitive characteristics of the DANTE pulse sequence in-depth and overcame a series of problems caused by the combination DANTE and MATRIX.We optimized and improved the DANTEMATRIX combination sequence to carotid artery wall imaging with a high signal-tonoise ratio and contrast.Secondly,we solved rapid vessel wall image reconstruction to shorten the clinical scan time for blood vessel wall inspection.The rapid imaging technology combining wave-controlled aliasing in parallel imaging and compressed sensing（wave-CAIPI-CS）was first proposed.In addition,we employed the deep learning reconstruction technology to overcome the technical limitations of current compressed sensing iterative reconstruction,breaking through the existing compressed sensing iterative reconstruction algorithm requiring parameters adjustment based on experience and bottleneck of low computational efficiency.We designed the network structure of multicontrast deep learning joint reconstruction and made use of information redundancy in higher dimensions of multi-contrast three-dimensional blood vessel wall imaging,to further improve the acceleration space and image reconstruction quality of multicontrast vessel wall imaging.Thirdly,we built an intelligent analysis method based on deep learning.The machine learning method was applied to segment 3D time-of-flight magnetic resonance angiography（TOF-MRA）to obtain craniometrical blood vessel information and extract the centerline of the blood vessel.It was further to match with the 3D blood vessel wall image,and the blood vessel wall image was reconstructed by a straightened surface.Straightening the curved vessel wall to display in a 2D plane is helpful for the doctors to observe the vessel wall and plaque shape intuitively and quickly.Moreover,this dissertation studied and developed an optimized deep learning network（D-UNet）structure to segment the 2D cross-sectional image of the target blood vessel perpendicular to the centerline,which could calculate local disease and quantitatively analyze the overall blood vessel.This method can realize the whole-process intelligent analysis of vascular plaque,which has been applied to domestic magnetic resonance imaging systems.