Research On Heart Segmentation Based On Deep Learning

Cardiovascular disease is one of the leading causes of death in the world,and its diagnosis and subsequent treatment relies on multiple cardiac imaging modalities,including contrastenhanced ultrasound,computed tomography,and magnetic resonance imaging.Among them,Cine-Cardiac magnetic resonance imaging(Cine-CMRI)is recognized as a template for noninvasive assessment of cardiac function,and with appropriate protocols,can also provide accurate information on morphology,tissue viability or blood flow.It is the gold standard for assessing cardiac function.However,accurately segmenting it is a challenging task,due to the many artifacts in cardiac images,large gaps in target organs,low short-axis resolution,and high cost of label data acquisition.Therefore,the current heart segmentation method is still dominated by manual segmentation.This paper uses deep learning convolutional neural network to segment cardiac dynamic MRI images.According to the different image characteristics and task requirements,three cardiac segmentation algorithms are designed:(1)A heart segmentation model based on dense residual coding blocks and fusion of multiscale features is proposed.Given that the cardiac dynamic MRI images belong to thick slice data,we chose to preprocess the 3D images into 2D slices and then perform segmentation.The model uses Unet++ as the basic model,and improves the encoding block and decoding block into dense residual encoding block and global attention decoding block respectively.The combination of Unet++ and dense residual coding block can make full use of the information on the 2D slice,in which Unet++ provides the network with decoding information of different depths and different scales.And the dense residual coding block can effectively aggregate deep coding features and shallow coding features,at the same time,it can shortens the distance between the coding layer and the coding layer.The global attention decoding block provides the network with global information of the feature map during the decoding process,so that the decoding features can be effectively aggregated,and the restoration results are more in line with the physical characteristics between the cardiac sub-results.(2)A heart segmentation model based on feature enhancement and feature correction is proposed.The network is a multi-input,multi-branch,multi-task 2.5D segmentation network that uses temporal information for feature enhancement and spatial information for feature correction.First of all,in order to make full use of the temporal information of dynamic cardiac MRI images to compensate for the sparse 3D context information,a new temporal enhancement coding module is proposed.The target frame that needs to be segmented and a continuous time segment containing the target frame are input together as a key sequence.The key sequence is used to obtain richer temporal features,and the target frame can provide more accurate edge features.Secondly,in order to aggregate more beneficial features and better fuse temporal and edge information,a deformable global connection is proposed to provide a wider range of multidimensional feature information for the decoding part of the network.Finally,the feature correction is performed on the original segmentation results by using the additional spatial direction field learned in the training process.(3)A semi-supervised cardiac segmentation network based on the combination of selflearning and cross-teaching is proposed.In order to alleviate the pressure of obtaining labeled data of Cine-CMRI and use a small amount of labeled data and a large amount of unlabeled data to train the segmentation network,a semi-supervised heart segmentation model based on Unet self-learning and cross-teaching between Unet and Transformer is proposed.This model uses Unet as the main model,and the Transformer and Unet network itself as auxiliary models.First of all,in the self-study stage,an explicit consistent supervision method is adopted,and it assumes the two identities of teacher and student at the same time.As a teacher,it generates learning goals for students.As a student,it conducts supervised learning with a small amount of labeled data and learning objectives generated by the teacher model.In the cross-teaching stage,in order to make up for the defect that Unet cannot obtain long-distance information,the Transformer network is introduced,and the implicit consistency supervision method is adopted.Unet and Transformer are pseudo-labels for each other,which directly realizes end-to-end mutual supervision.In this process,the local feature extraction ability of Unet and the longrange relationship modeling ability of Transformer can be effectively combined,and the different sensory attributes of the two networks can be used to complement each other to obtain better segmentation results.