Studies On Magnetic Resonance Imaging Reconstruction And Computer-aided Diagnosis Based On Artificial Intelligence

Magnetic Resonance Imaging(MRI)is an important clinical imaging modality.Different MRI sequences can be used to provide different information about body structure,function,and metabolism.MRI has the unique advantages of high contrast,high resolution,arbitrary orientation,and imposes no radiation damage to patients.While MRI plays a unreplacable role in clinical environment,it is also one of the most complex imaging modality.A lot of scientific and technical issues need to be addressed in data acquisition,image reconstruction,image post-processing,and image analysis.In this dissertation,the data-driven and artificial intelligence(AI)methods are used to study MRI reconstruction,processing,and analysis.These studies are partitioned into two parts.1.We focused on the high-resolution MRI reconstruction.High-resolution MR images requires long scanning time,which increases the cost of MRI examination.MRI reconstruction from partial k-space data(under-sampling)is an important method to shorten the scanning time.Firstly,we studied the MR angiography(MRA)reconstruction.Based on the fact that the full width of the half maximum(FWHM)of the objects in the image is resolution-invariant,we proposed an iterative method called COnstrained Data Exploration(CODE).CODE utilized the low-frequency k-space data to reconstruct high-resolution MRA images.We compared CODE with compressed sensing(CS)reconstruction on both simulated and real data.CODE produced images with sharper edges and produced better evaluation of the stenosis of the carotid vessels.Secondly,we incorporated image segmentation into dictionary learning(DL)algorithm and proposed a novel method called DL with segmentation(DLS).DLS utilized the data-driven structures of DL and the a priori knowledge of MR brain images to reconstruct under-sampled brain images.Experiments on simulated and in-vivo MRI data proved DLS can achieve better quality than the DL reconstruction.2.We aimed to develop a computer-aided diagnosis(CAD)system for prostate cancer(PCa)diagnosis based on multi-parametric MRI(mp-MRI).We carried out a series of studies on prostate segmentation,PCa detection,and PCa diagnosis.While Mp-MR provides richer diagnositic information,it also takes more time to interpret.CAD system analyzes images and provides useful suggestions and hints to help radiologists to make diagnostic desicions.Thus,it can relieve the burden of the radiologists,increase the working efficiency and decrease the misdiagnosis rate.To develop a prostate cancer CAD system,we first studied prostate segmentation.We used a convolutional neural network(CNN)based on U-Net,a classic configuration of CNN,with multi-slice input and multi-supervision output.This network acquired a good segmentation result,especially for the apex and base.Then based on this segmentation result,we proposed the TrumpetNet using mp-MRI as input for PCa detection.A multicenter study proved TrumpetNet achieved high sensitivity and could be helpful to radiologists.We also developed a radiomics model to diffentiate clinical significant PCa from non-clinical significant PCa,and a deep learning model to classify the PCa and non-PCa.We discussed the effect of the different combinations of MRI sequences and structures of the network.We also proposed a novel prediction method,namely enhanced prediction,to increase the predictive ability of the model.During the above studies,we developed two software.FeAture Explorer(FAE)can be used in radiomics studies.FAE could be used to explore the different combinations of methods and found the best one.FAE has been used in several hospitals now.DeepOncoAnalysis can be used to label MR images,extract features,and preprocess data.We also integrated above-mentioned PCa models in DeepOncoAnalysis so that they can be used for clinical research.It turned out to be a convenient tool for collecting feedbacks from the radiologist for model refining.