Automatic Segmentation Methods Of MRI Cardiac Images Based On FCN

Cardiovascular diseases pose a serious threat to human life and health.This situation is even more severe in China.At present,the effective diagnosis and treatment of cardiovascular diseases is cardiac MRI imaging which annotated by professional radiological doctors to calculating the various functional parameters of the heart to analyze the patient's condition.However,the demand for manual labeling is still expanding,and the labeling process is also cumbersome and time-consuming.There are also differences in labeling between different doctors.Deep learning has achieved great success in the field of semantic segmentation.There have been some attempts to use deep learning to implement medical image segmentation tasks,but there is still the possibility of further improvement for MRI image automatic heart segmentation tasks.In this paper,using deep learning FCN technology,for the heart segmentation task of two-dimensional and three-dimensional data,respectively,the MRI image heart automatic segmentation method is given.For two-dimensional DCM format left ventricular contour segmentation task,this paper proposed a convolutional neural network based on group normalization and nearest neighbor interpolation upsampling which we called GNINU U-net(U-net with Group Normalization and Inter Nearest Upsampling).We constructed a convolution module with group normalization operation based on group normalization method for fast and accurately feature extraction.Based on nearest neighbor interpolation method,we constructed an up-sampling module for feature restoring.We conducted a pre-processing method for Center cropping ROI extraction and a detailed controlled experiment of GNNI U-net on the Sunnybrook left ventricular segmentation dataset and the LVSC left ventricular segmentation dataset.The experimental results show that the GNNI U-net obtained the accuracy of Dice coefficient of 0.937 and the accuracy of Jaccard coefficient of 0.893 on the Sunnybrook dataset,it obtained the accuracy of the Dice coefficient of 0.957 and the accuracy of the Jaccard coefficient of 0.921 on the LVSC dataset.The GNNI U-net network achieves higher Dice coefficient accuracy than other convolutional network segmentation methods in the field of left ventricular contour segmentation.Finally,according to the experimental results we discussed and verified that the convolutional module of the normalization operation can accelerate the convergence of the network and improve the accuracy,and the up-sampling module using the nearest neighbor interpolation method is more friendly to the smaller target segmentation such as the left ventricle contour,and our model can accelerate network convergence to a certain extent.For the task of three-dimensional N? format MRI image total heart segmentation,this paper proposes 3D GNNI U-net based on GNNI U-net.In the ACDC data set,we conducted a pre-processing method of slice size dimension ROI extraction,slice number dimension nearest neighbor interpolation amplification and three-dimensional geometric transformation data enhancement method and a detailed controlled experiment of 3D GNNI U-net.The results show that the 3D GNNI U-net has achieved a certain accuracy in the Dice coefficient and Hausdorff distance,and is significantly improved compared to 3D U-net.This paper verifies that the 3D GNNI U-net maintains the good characteristics brought by the 2D GNNI U-net structure when processing three-dimensional data and the convolution module of the three-dimensional group normalization operation and the up-sampling module of the nearest neighbor interpolation have the function of accelerating convergence and improving the segmentation accuracy like when processing two-dimensional data.By analysis and discussion,this paper proofs that the 3D GNNI U-net maintains the good characteristics brought by the 2D GNNI U-net structure when facing the task of 3D data and multi-object segmentation.It is proofed that the convolution module based 3D group normalization operation and The up-sampling module of the neighbor interpolation method still accelerates the convergence of the network and improves the segmentation accuracy.Based on the FCN deep learning technology,this paper proposed the automatic heart segmentation method based on FCN for the automatic heart segmentation tasks of 2D and 3D MRI images.It is proved through experiments that the methods proposed in this paper have achieved excellent segmentation results,which has good reference value and practical significance for the scientific research work and clinical application of MRI image automatic heart segmentation.