Research On Segmentation Of Cardiac MRI Based On Deep Learning

As the power source of the human circulatory system,the heart is a complex and integrated system that combines electrophysiology,kinetics,blood fluid mechanics,and neurological and biochemical control.According to the data released by the World Health Organization,heart disease is the number one cause of death worldwide.In-depth study of heart anatomy,movement and other characteristics is of extraordinary significance for the prevention and diagnosis of heart diseases.In practice,it is necessary to rely on physicians to outline the heart in MRI images of the heart.It is a time-consuming task to segment the heart outline by manual sketching,and traditional methods are poor in segmentation accuracy in cardiac MRI image segmentation tasks,which cannot assist doctors’ work well.With the rapid development of deep learning,the application of deep learning-based methods to cardiac MRI image segmentation tasks has achieved impressive results.In this paper,we carry out a deep learning-based cardiac segmentation task,which consists of two main aspects as follows.Firstly,the left ventricular segmentation task is carried out using feature reuse approach.The feature reuse is enhanced by using dense blocks in the two main paths.An additional path is added to reduce the semantic divergence between the two main paths,and compressed dense blocks simplify the connectivity in the dense blocks,while a 1x1 convolution is used to compress the feature map at the end of the module.Using this approach prevents the model from having feature map explosion problems in the upsampling paths.In this paper,the improved method is tested on three left ventricle datasets,and the experimental results demonstrate the superiority of using feature reuse in doing ventricular segmentation while our method is convincing compared to some mainstream algorithms.Next,the biventricular segmentation task was further carried out.The dense block using feature reuse considers only the first-order statistical information features of the data.And the higher order statistical information of the data is ignored.In this paper,we try to use the second-order statistical information of the data to assist the model segmentation,and also use the multi-scale module to expand the information of the feature maps of the model.The feature maps of different scales are utilized in the upsampling path,and the final output of the path is cascaded with the feature maps of different scales in the path to achieve the goal of expanding the semantic information by increasing the number of feature maps.The attention mechanism can help the model to better focus on the region of interest and accelerate the optimization of the model while improving the model performance.We combine the higher-order statistical information with the attention mechanism,and consider both spatial attention and channel attention dimensions to suppress or reinforce different parts of the feature map.The experiments show that the multiscale module and the second-order statistical information of the hybrid attention module improve the performance of the model,and the experimental data demonstrate the performance improvement of the method in the dataset and the feasibility of the method by comparing some mainstream algorithms.