| According to the report from the World Health Organization(WHO),cardiac-related disease is one of the leading causes of death worldwide.Approximately 17.9 million people died from cardiac-related disease every year,among which heart failure is one of the main causes of death in patients with heart diseases,and the number of deaths is increasing year by year.In recent decades,significant progress has been made in the research and clinical practice of cardiac-related diseases,which have greatly improved the diagnosis and treatment of cardiac-related diseases and reduced their mortality.Nowadays,increasingly mature medical imaging technology continues to promote the development of clinical diagnosis technology,and medical imaging technology has been widely used in clinical examinations.Modern medical imaging techniques such as magnetic resonance imaging(MRI),computer tomography(CT)and ultrasound(US)can perform noninvasive qualitative and quantitative assessment of the anatomical structure and function of the heart for diagnosis,disease monitoring,and treatment Plan and prognosis.Compared with static imaging technology,dynamic medical imaging technology,especially dynamic cardiac CT imaging(4DCT),can capture the morphological changes of the heart with high temporal and spatial resolution during the entire motion cycle of the heart,and assist clinical evaluation of cardiac static and dynamic parameters,detect and locate the local or small structural lesions of the cardiac.This is an important auxiliary method for early diagnosis,precise positioning,treatment planformulation,and prognosis evaluation of cardiac disease.At present,the clinical analysis for cardiac dynamic evaluation highly rely in the dynamic imaging examination,especially 4D-CT images.However,the tremendous dynamic image data has brought huge challenges to clinical analysis.Traditional manual annotation-based analysis requires a lot of manpower,and manual annotation results are difficult to ensure timing consistency and easy to introduce human errors.Therefore,there is a great requirement for efficient and accurate automatic dynamic image analysis algorithms to relieve the pressure of data analysis and processing.For clinical analysis of cardiac dynamic images,image segmentation and motion estimation of target organs are important steps for quantitative analysis of cardiac image dynamics.Among them,image segmentation is an important first step in many applications.It processes the image into multiple semantically(i.e.,anatomically)meaningful regions based on which quantitative measurements such as myocardial mass,wall thickness,left ventricular(LV)volume,and ejection fraction(EF)can be extracted.Subsequently,motion estimation based on image registration can track the movement of the heart during the cardiac cycle,provide quantitative analysis indicators for dynamic assessment of the heart state,and then perform quantitative analysis and precise positioning of abnormal motion areas.However,existing medical image processing methods are mostly based on the calculation of local features of the image to process and analyze the target,which is usually limited to its performance,generalization ability,etc.,and cannot meet clinical requirements.Moreover,for motion estimation,the existing algorithms usually estimate the motion deformation by calculating the similarity of paired-image local features,which are better for local deformation,but cannot accurately estimate large morphological and structural deformations in the cardiac cycle.In this study,we mainly focus on the analysis and processing of 4D-CT cardiac images,and propose several relevant processing algorithms to address the challenges in dynamic medical image processing.The main innovations of this article is following:· For the continuity of 4D-CT images in time series and the challenge of segmentation at the end of systole,a spatial sequence network(SSNet)is proposed to obtain the deformation and motion characteristics of LVC with an unsupervised learning model,and combine the twoway motion deformation information with the image Contextual information integration fusion learning,assisting dynamic left ventricular segmentation,improving the consistency and accuracy of segmentation in time series.· In view of the huge changes in heart shape and irregularities in4D-CT dynamic images,this study provides a dense-sparse-dense(Dense-Sparse-Dense,DSD)unsupervised motion estimation network,which extracts the original dense image The sparse flag is used to represent the topological structure of the target organ to discard redundant information and be used for the motion estimation of the target organ.Subsequently,the sparsely expressed displacement reconstruction is mapped back to the dense image domain through the motion reconstruction network to construct the motion field,which is used to improve the quality of motion estimation and reduce the wrong motion estimation.· In order to increase the time sampling rate of 4D-CT images and reduce the radiation damage caused by high-dose CT scans to patients,a dynamic medical image spatiotemporal volumetric interpolation network(SVIN)is proposed.SVIN introduced a dual network learning strategy: spatiotemporal deformation estimation network and two-way interpolation optimization network.First,the deformation estimation network uses 3D convolutional neural network(CNN)for unsupervised parameter registration,and derives the motion deformation field from a pair of images.Then the bidirectional interpolation optimization network uses the derived motion field to interpolate the image volume.· According to the proposed automatic segmentation algorithm and motion estimation algorithm,the left ventricular surface motion vector can be extracted,and the motion vector features,such as motion amplitude,direction,etc.,can be used for cluster analysis to locate the abnormal motion area to assist clinical diagnosis and precise preoperative positioning.In summary,this study successfully employs the deep learning technology into dynamic medical image processing,and effectively addresses the challenges in dynamic cardiac image segmentation and motion estimation.Specifically,this study presents several deep learning based approaches,including a spatial-sequential segmentation model for dynamic left ventricular segmentation,an unsupervised landmark detection guided motion estimation model,a motion-guided dynamic image interpolation approach,and a cluster analysis method based on motion vectors,respectively.These proposed algorithms can be used to assist clinical analysis and diagnosis of 4D-CT image,and provides a reliable basis for the practical application of computeraided technology in dynamic image analysis. |
PDF Full Text Request