Research On Left Ventricle Image Quantification Method Based On Deep Multi-task Network

Automatic quantitative analysis of Magnetic Resonance(MR)images are of great clinical significance for early detection and diagnosis of cardiac diseases.In early clinical practice,the quantification of cardiac MR images was obtained manually,that is,by calculating the boundary of the heart contour manually drawn by clinicians to obtain relevant clinical index.However,in the face of massive medical image data,this artificial method is time-consuming and cannot guarantee the accuracy of the results.Therefore,automatic algorithms of quantification of cardiac MR images are urgently needed.Moreover,due to the complex structure,uneven gray distribution and blurred boundary of the cardiac MR image,the quantification of the images is faced with great challenges.It is a popular trend to adopt end-to-end deep learning technology to directly implement image quantification of cardiac MR images.This paper has proposed a deep neural network model algorithm based on deep learning technology for automatic quantification of cardiac MR images.Relevant research contents are as follows:An automatic quantification algorithm of left ventricle(LV)MR images based on the deep multi-task network.In terms of innovation,we utilize the segmentation results to optimize the original cardiac MR image data.Compared with the original image,the segmented binary contour image discards the background interference information of the original image.Therefore,the quantification results estimated by the deep neural network is more accurate.In terms of network design,first,we have designed a segmentation network for image segmentation,which adopts the current popular encoder/decoder structure.On the one hand,it can sample the input original image and generate a feature map with low resolution,on the other hand,it can restore the feature map to a full-resolution feature map through up-sampling.After that,the segmented binary contour image is taken as the input of the quantification network,and more effective image features obtained by fusing the output of multiple convolutional neural networks(CNN)models.Finally,combining with the long short-term memory network(LSTM)to capture the dynamic deformation information of the MR sequence image.This framework we designed takes into account the connection between image segmentation and quantitative analysis,which not only realizes automatic image segmentation but also guides the process of image quantitative analysis through the result of image segmentation,making the automatic quantification results of heart MR images more accurate.Based on the above proposed automatic segmentation and quantification method,this paper has realized accurate automatic segmentation and quantification analysis of cardiac MR images.The results of the experiments on 2,900 heart MR images from 145 patients show that our proposed framework(Indices-JSQ)has excellent performance in both automatic image segmentation and quantitative analysis.In terms of segmentation performance,the Dice metric(DM)score is 0.87.In the respect of quantitative performance,the mean absolute error(MAE)of myocardium area(A-myo)and cavity area(A-cav)is 157 mm~2,the MAE score of the dimension of the cardiac cavity in three different directions(dim1,dim2,dim3)is 2.43 mm,the MAE score of regional wall thicknesses in six different directions(RWTs)is 1.29 mm.The experimental results show that Indices-JSQ will make the early detection and diagnosis of heart disease more efficient and convenient for clinical workers,and have a broad clinical application prospect in the future.