Research On Key Technologies Of Automatic Classification And Segmentation Of Cardiac Data Based On Deep Learning

At present,the mortality rate of cardiovascular diseases is the main cause of death,which has been a common threat to all mankind.It is significant to explore a fully automatic classification and segmentation algorithm for the clinical prevention and diagnosis of cardiovascular diseases.Meanwhile,it also helps physicians free themselves from tedious and time-consuming handcrafted classification and segmentation tasks,which has great practical significance.The electrocardiogram contains a wealth of information about the activity of the human heart,which reflects the heart health of physiological activities.It is one of the key factors for evaluating heart function and determining heart disease.And cardiac magnetic resonance imaging is a type of non-invasive cardiac imaging technology and an important method for diagnosing heart and macrovascular diseases.It has become an important gist for the detection and evaluation of cardiac structure and function.Our topic will focus on cardiac data based on deep learning theory,which mainly to study the classification task of arrhythmia using one-dimensional ECG signals and cardiac images segmentation task using two-dimensional data of cardiac MRI.This paper mainly contains the following research contents and achievement:Firstly,we study the preprocessing technology of ECG signals.The research direction of preprocessing is mainly on denoising.A denoising method based on finer morphological wavelet transform theory is provided,called the Improved Morphology-WT model.On the one hand,the model analyzes the feasibility of morphological filtering and wavelet transformation methods theoretically.And wavelet basis functions,decomposition levels,threshold processing methods,and the improved threshold estimation functions also are determined.On the other hand,a large number of experiments have verified the process of determining wavelet basis functions and decomposition levels.The final experimental results also show that the Improved Morphology-WT method is very effective in removing low-frequency and high-frequency noises in the ECG signal,and it lays the foundation for the following research work.Secondly,the CB-LinkNet model based on a convolutional neural network and bidirectional long and short-term memory model is proposed to detect and classify atrial fibrillation signals in this paper.The convolutional neural network has powerful feature extraction capabilities,but it is proposed to mainly solve image classification,object detection,and image segmentation.And it does not focus on a one-dimensional signal.The ECG signal is essentially time series data.A bidirectional long and short-term memory network is used to supplement and adjust the learning ability of the convolutional neural network,making the network model more suitable for the feature extraction task of time-series signals.Meanwhile,the paper obtains two input signals,RR interval data(set A)and heartbeat sequence data(set B)based on the MIT-BIH Atrial Fibrillation database.They will be leveraged to verify sensitivity of atrial fibrillation signal to data.The three sets of ablation experiments designed in the paper finally fully verify the robustness and efficiency of the model.The final classification accuracy reaches 99.94%and 98.63%in the training and validation stages,and 96.59%on the test set.The sensitivity and specificity reach 99.93%and 97.03%on the test set,respectively.Meanwhile,this paper compares the recent models and results of atrial fibrillation detection and fully analyzes the practical value of this research.Thirdly,this paper proposes three deep neural classification network models,including a plain-CNN model and two MSF-CNN models(A and B)are proposed for the detection and classification of multiple types of arrhythmia signals.The plain-CNN model is a baseline network structure with multiple convolutional layers;MSF-CNN A is proposed based on the plain-CNN model to improve the learning ability of the plain-CNN network,mainly by adding parallel-group convolutional block(including three different convolution kernels,namely 1×7,1×5,1×3).And then we provide the MSF-CNN B model based on the MSF-CNN A through the implementation of parallel and concatenation-group convolution block and residual learning blocks.In terms of the features of arrhythmia signals,this paper designs a multi-scale input signal for the model to validate the impact of data scale on the model's performance.At the same time,the data augmentation methods are leveraged in one-dimensional signals to improve the scientific and validity of data.Six sets of ablation experiments prove the generalization ability and robustness of the model,and finally achieved the accuracy of 96.59%,the sensitivity of 99.93%,and the specificity of 97.03%on the test set,which fully reflects the model's important value to the classification tasks of arrhythmia signals.Finally,this paper mainly proposes the Res-LinkNet segmentation model for the segmentation of left ventricular MRI images.It is divided into three parts:encoder,center,and decoder.The core part of encoder is that the model is ResLink block,this block replaces the ResNet block in D-LinkNet.The ResNet block is designed for classification tasks.The design limits the size of the receptive field and lacks cross-link feature interaction.The ResLink block can perfectly solve this problem.The main structure of the center part is DenseASPP.DenseASPP is mainly to effectively solve the problem of the receptive field of the feature map.It includes a baseline network,which is connected to 5 dilated convolutions to achieve hybrid dilated convolution.And the dilatation rates are 3,6,12,18,and 24,respectively.On the one hand,this structure increases the receptive field in the deep layers.On the other hand,it effectively alleviates the grid problem caused by ordinary dilated convolution operations.The features from different channels are also fused.The decoder part mainly uses Dense up-sampling convolution operations to enlarge feature maps to the required size.This mainly is to restore the image feature information to the highest extent.The experiment includes three sets of ablation analysis,and the results show that Res-LinkNet101 reached the mean pixel accuracy of 99.88%,the mIOU of 94.95%,and the F1 score of 95.57%,which fully reflects the effectiveness of the model.In a word,the main innovations of this paper are as follows:(1)This paper proposes an improved Morphology-WT denoising model to process the ECG signals.And an adaptive threshold estimation method and an improved threshold function are also proposed to complete the wavelet decomposition in the denoising model.(2)A CB-Linknet model that integrates the convolutional neural network,and the bidirectional long-term and short-term memory model is proposed to detect and classify the Atrial fibrillation signals.In the experiment,we design two types of input signals to evaluate the characteristic sensitivity of atrial fibrillation signals.(3)In the study on the detection and classification of multiple arrhythmia signals,this paper proposes three end-to-end classification network models,and innovatively designs data augmentation methods in the signals to effectively avoid overfitting.(4)In the study of left ventricle segmentation,the paper proposes Res-Linknet segmentation model,and design a ResLink structure model based on attention to complete the feature extraction of cardiac magnetic images in the encoder part.