Segmentation And Analysis Of Cardiac Magnetic Resonance Image Based On Deep Learning

Cardiac function analysis plays an important role in clinical cardiology for patient management,disease diagnosis,risk assessment and treatment decisions.Cardiac Magnetic Resonance(CMR)Images is mainly used for cardiac function assessment and cardiovascular diseases diagnosis.CMR images is considered to be the gold standard for non-invasive cardiac function analysis and estimation of clinical parameters.Depicting important organs and structures from medical images is often the main step in estimating clinical parameters,disease diagnosis,prognosis prediction,and surgical planning.In clinical diagnosis,radiologists manually or semi-automatically draw outlines containing the left ventricle(LV),right ventricle(RV),and myocardium(MYO)of the heart to obtain clinical parameters related to cardiac function analysis(such as ejection fraction,stroke volume,etc.).However,manual or semi-automatic segmentation is often time-consuming and easy to introduce manual errors.On the one hand,it brings a huge workload to the doctor,on the other hand,it is often affected by factors such as the doctor’s diagnosis experience and inconsistencies.In recent years,the incidence of heart-related diseases has gradually increased,and manual or semi-automatic segmentation is difficult to meet the huge medical needs.Therefore,there is an urgent need for a fast and accurate automated CMR image segmentation algorithm to improve diagnostic efficiency.In this paper,we first design a network named Cardiac Segmentation Network(CSNet)suitable for segmentation of multiple cardiac structures(left ventricle,right ventricle,and myocardium)based on fully convolutional neural networks(FCN).CSNet divides the CMR image segmentation task into two stages.The first stage sets up a 2D U-Net and a parallel 3D U-Net,which are used for two-dimensional shape features and three-dimensional structural features learning;the second stage sets up a 2D U-Net with fused features as input is used to obtain the final segmentation result.The fusion feature is obtained by weighted addition of the output feature maps of the two networks in the first stage through a learnable weighted addition operation.Its main function is to introduce 3D structural information for the 2D U-Net in the second stage to compensate for the 2D U-Net can not learn the defects of three-dimensional structural features,thereby improving model performance.Experimental results based on the ACDC data set show that CSNet achieves the best performance compared with 2D U-Net and 3D U-Net.CSNet achieves good performance,but introduces a large amount of parameters and calculations.In order to optimize the amounts of parameters and calculations as much as possible while ensuring the performance of the model,this paper further improved CSNet.The improved network(Cardiac Segmentation and Deep Feature Fusion Network,CS-DFFNet)embeds the 3D convolution kernel into the 2D convolutional neural network,combines the output features of 2D convolution and 3D convolution more deeply,and integrates the self-designed Multi-Structure Coarse-to-Fine Inference Module(MSCFIM)and loss function based on volume information.The introduction of 3D convolution can make up for the defect that 2D convolution cannot learn the inter-plane features of 3D CMR images.Besides,the MSCFIM takes the cardiac multi-structure consistency and continuity into consideration,and introduces more sufficient global and local information to enhance the network ability of better recognization of the fine structures in CMR images.The loss function based on volume information uses the volume values of the 3D images to be segmented as a constraint term,which speeds up the network training process and brings a certain performance improvement.Compared with typical 2D FCN,CS-DFFNet can take both 2D shape and 3D structure features of 3D CMRI into account,and making up for the shortcomings of 2D FCN which cannot learn 3D structure features.In addition,large memory consumption and over-fitting problems of the original 3D FCN are improved.Experiments showed that compared with CSNet,CS-DFFNet has been improved in performance,and the amount of parameters and calculation is much less than CSNet.Experiments show that CS-DFFNet achieved better performance but contained less parameters and calculation than CSNet.The test results based on the ACDC test set show that the average dice coefficient of the multi-structure segmentation results of the LV,RV,and MYO of the method proposed in this study is 0.945,0.925,0.91 respectively,some of the indicators surpass the algorithm of the ACDC data set that has been ranked first in the segmentation accuracy.In addition,a heart disease classification algorithm based on the results of cardiac magnetic resonance image segmentation achieves(five different disease types image data)94% classification accuracy.