Orientation Of The Left Ventricular Long Axis Based On Deep Learning

With the increasing clinical application of magnetic resonance imaging,cardiac magnetic resonance imaging has become the "gold standard" for judging the structure and function of the heart.It has the characteristics of large field of view,high tissue resolution and no radiation.In the past,due to the motility of the heart and the inconsistency between the long axis of the heart and the long axis of the human body,one of the biggest difficulties in scanning the heart was positioning.The whole process was manually performed by a professional.During the heart scan process,in order to complete several basic level selections,we need to scan once and locate once.In the past 20 years,a large number of methods for heart segmentation have been proposed,including semi-automatic and automatic segmentation.However,there is very little work related to automatic planning of cardiac scans.It is still a challenge in clinical practice today.Currently,related technologies have been used for commercial use,but they are security technologies.Therefore,in order to automatic orientate the long axis,the relevant algorithms must be researched independently.This paper proposes an automatic localizing method for the left ventricular long axis using deep learning.Main tasks as follows:(1)The current short-axis orientation related method is studied.The biggest drawback is that the process is too complicated and relies too much on the previous feature extraction work.Therefore,this article uses deep learning method.A variety of neural networks have been studied,including convolutional neural networks and recurrent neural networks,as well as various modules.The localizing method of the long axis is constructed in combination with related technologies.(2)Because the data is not enough,this paper simulates the real pre-scan 3D data by using the existing 3D data,and studies the method of estimating the left ventricular long axis orientation by 3D volume magnetic resonance data and 2D transverse slices.(3)Five models are proposed,two for 3D data and three for 2D data.Due to the large amount of computation of the 3D model,this paper proposes a method for reducing dimensionality using the channel attention mechanism and a 3D model based on Inception.The first model of 2D data uses a convolutional neural network to extract planar features,and a recurrent neural network to extract spatial features.The second and the third model use continuous transverse slices as input,using DenseNet and Attention mechanism for long-axis positioning.(4)In order to improve the accuracy of positioning.Summarize the skills of selecting hyperparameters,activation functions and optimization functions in different model training processes.(5)The positioning effects of the five models were tested,and the advantages and disadvantages of the model were analyzed from the aspects of absolute error mean,standard deviation and time.First,the simulation data was used for the experiment.The results showed that the 3D_Attention model had the best result.The α test result was 3.18±4.07°(mean±variance),and the β test result was 2.57±3.14°.The training time is also short,the average test time is 20.5 milliseconds.Then 11 real data tests are used,the results show that 2D_DenseNet works best,the α test result is 5.36±6.24°,and the β test result is 6.82±4.65°.