The Research Of Supervised And Cross-modal Cardiac Segmentation Based On The UNet Network

Cardiovascular disease is an important public health problem,and its mortality rate ranks first among various diseases.It is of great significance to prevent and treat cardiovascular disease.In clinical practice,quantitative analysis of cardiac anatomy is helpful for the diagnosis and treatment of various cardiovascular diseases.Manual cardiac segmentation is a highly professional,difficult and time-consuming task.So,completing this step efficiently and fully automatically will reduce the burden on doctors and the waiting time of patients,which is of great significance for the diagnosis and treatment of cardiovascular diseases.With the application and development of computer vision technology in various fields,the research of artificial intelligence-assisted medical treatment in the medical scene has also made great progress.Fully automatic cardiac segmentation based on deep learning has also become a research hotspot.In practice,researchers have established large cardiac datasets,hoping to obtain better cardiac segmentation results through supervised training,in order to meet the high-precision requirements of clinical diagnosis.However,there are many challenges in the cardiac segmentation task,for example,the medical scenes in reality are very complex,cardiac shape,size,and lesions are diverse,and there are discontinuous areas in the cardiac image that are easy to confuse the boundary.It is difficult to obtain professional ground truth of segmentation and so on.In order to solve the above problem,considering the characteristics of discontinuity and multi-modality of medical images,we propose the research of supervised and cross-modal cardiac segmentation based on the UNet network,which is dedicated to optimizing the segmentation boundary and alleviating the difficulty of obtaining cardiac segmentation ground truth.In order to obtain more accurate boundaries of cardiac segmentation,we propose a supervised cardiac segmentation that pays more attention to discontinuity.Based on the medical image features of cardiac magnetic resonance imaging(MRI)modality 2D images,it mines the discontinuous regions that are easily confused as boundaries,and pay more attention to the segmentation of discontinuous regions during the training process to optimize the segmentation results.The experimental results on the self-collected large cardiac segmentation dataset show that compared with other algorithms based on cardiac contour and boundary,our algorithm has achieved better experimental performance than existing methods.In order to alleviate the difficulty of obtaining the ground truth,we propose a twostage cross-modal cardiac segmentation algorithm based on adversarial training and online pseudo-labeling,which is realized unsupervised cross-modal cardiac segmentation.The cardiac data in different modalities in the clinic have the same cardiac structure and semantic information but different image features.In the first stage,we use the cycle-consistent adversarial networks to align the appearance of different modal images.In the second stage,in addition to further aligning the entropy space of different output data,we also generate online pseudo-labels by Mean Teacher architecture for the target modal,to assist segmentation training of target modal.Compared with the method which continuous stacking of adversarial training modules to optimize the performance of the target modal segmentation result,the effect of pseudo-labeling technology is more significant when the difference between the two modalities is small.We conducted experiments on the public MM-WHS cardiac segmentation dataset,which realized two-direction cross-modal cardiac segmentation of cardiac Magnetic Resonance Imaging(MRI)and Computed Tomography(CT)modalities,and both obtained the current best performance.In this paper,we focus on the cardiac segmentation task in medical image analysis.Thanks to considering the characteristics of discontinuity and multi-modality of medical images,we optimized the adverse effects of discontinuous areas in cardiac imaging,and alleviated the problem of difficulty in obtaining the label of cardiac segmentation by the cross-modal segmentation algorithm that combines adversarial training and pseudo-labels.