Research On Cardiac MRI Segmentation Based On Deep Learning

There are two main types of tasks in deep learning and its application in medical diagnosis.The first type of task is classification,which reduces potential misclassification by mapping data to a specific result classification.The second type of task is to identify,that is,to use medical images and other types of ancillary data to identify and diagnose tumors or other diseases.The work of this paper is to do more precise pixel-level segmentation after classification.In the diagnosis of congenital heart disease,the first thing need to create a specific heart model for children who has complex congenital heart defects(CHD),and the first step in creating a model is to split the blood pool and myocardium from 3D cardiovascular magnetic resonance(CMR)images.Good segmentation results can be used to reconstruct an accurate three-dimensional heart model for clinicians to perform a heart surgery.Manually labeling the myocardium and blood pool based on cardiac MR images is the most reliable method,but there always are hundreds of images in an individual's cardiac MR scan.Performing pixel-level annotation on so many slices is a high-intensity physical and mental work,and the results of the labeling are seriously affected by the subjective consciousness of the observer.For different study individuals,automatic and accurate segmentation of cardiac magnetic resonance images has become a problem due to differences in heart shape(especially right ventricle),inconsistent signal intensities,and differences in signal-to-noise ratios of acquired data.This paper draws on the application of Dense Net and U-Net in MRI segmentation,and combine current popular methods attention mechanism and dilation convolution to propose an improved network structure for blood pool and myocardial segmentation of cardiac magnetic resonance images,the main improvements are as follows:1?Sparse block model was proposed based on the network structure of Dense Net and U-Net2?In order to expand the receptive field of the feature map,the mixed dilation convolution is used in a 3D manner3?Spatial attention model and Scale attention model based on channel attention model are added to our networkThe improved method proposed in this paper has been fully tested on the HVSMR 2016 dataset.Compared with other methods,the accuracy rate has been improved,and it has positive practical significance for other segmentation tasks which uses different type MRI data.