Segmentation And Quantitative Analysis Of Ventricular Magnetic Resonance Images Based On Domain Knowledge

Heart diseases are main menaces of human health,and 80% of them come from ventricular dysfunction.Hence,ventricle quantification is important for diagnosis and treatment of cardiac diseases in clinical application.Magnetic resonance imaging(MRI),which has the highest resolution for the soft tissue currently,is considered as the gold standard for the ventricle quantification.However,accurate quantification of ventricular indices is still a challenging task in clinical application.Firstly,the accurate quantification of ventricular function relies on accurate label of each image.There are about 200 MR cardiac images for every patient.Secondly,the motion artifact of ventricle MR images,which is caused by heart movement,make it more difficult to label the ventricular MR images.Thirdly,quantification of ventricular function indices mainly relies on manual label of ventricular MRI.This process requires a lot of time and labor,besides there are some subjective differences.Hence,efficient and accurate methods for quantification of ventricular function have an important clinical value and scientific significance.In this thesis,the key technologies of automated methods for quantification of ventricular function are studied.The specific contents are as follows:Firstly,a ventricle detection method was proposed,based on the prior knowledge in the field of ventricle MRI.The proposed method combine the intersection of space and the multiscale atlas matching method to address ventricle region detection.It is obviously superior compared with the state-of-the-art ventricle detection methods,and can efficiently detect the key point,i.e.,the center point of ventricle cavity.The proposed method had been demonstrated on the large-scale baseline dataset of ventricle MR image.This method is the base of the following key technologies of ventricle quantification.Secondly,based on the effective detection of ventricle,ventricle segmentation method was studied.Considering the imaging mechanism of MRI,ventricle's motion mechanism and geometric structure,based on the prior knowledge in the field of ventricle MRI,a new ventricle segmentation method based on spatial-temporal area change correlation is proposed.The method effectively modelled the area change correlation among different chambers,and decomposed the area change correlation into spatial area change correlation and temporal area change correlation.Compared with other ventricle segmen-tation method,the proposed method achieved more accurate segmentation,which had been proven by the experiment results on several baseline ventricle segmentation datasets.Based on the accurate segmentation of ventricle,the accurate ventricle quantification can be achieved.Thirdly,except for ventricle segmentation method,direct index regression method is another key technology for ventricle quantification.Hence,a ventricle index regression method also was studied.The experiments for index prediction of ventricle function were done based on the input with multi-view slices.The experiment results showed that the combination of slices on different view has an important impact on the accuracy of prediction results.Based on this discovery,the feasibility of combination strategy of multi-view slices was demonstrated by a large number of experiments.Hence,an optimal combination strategy of multi-view slices was proposed.Based on the strategy,an error correction strategy based on dynamic evolutionary network and ejection fraction correlation method was proposed.The proposed method not only can regress the indices of ventricular function,but also can improve the accuracy and efficiency of quantification of ventricular indices.Compared with the ventricle segmentation method for quantification of ventricular indices,the proposed method can directly get the final indices of ventricle,avoiding the reduction of accuracy and efficiency from multi-stage computing.Finally,based on the above research,a unified venticle quantification method was proposed.The proposed method combines the ventricle segmentation method and regression method into a unified framework.This method utlizes the output consistency of two ventricle quantification tasks to construct a correlation constraint for loss fucntion.Then,based on the output consistency of indices,an automated uncertainty assessment strategy for the output of ventricle index was proposed.This strategy improved the accuracy of the quantification of ventricular indices.The proposed method in this part not only can accurately quantify the ventricular indicators,but also can effectively achieve uncertainty assessment.The high clinical application value has been demonstrated on several baseline datasets of ventricle quantification.