Research On Medical Signal Recognition Method Based On Residual Network

For diseases such as cancer and heart disease,which have very high mortality rates.Early diagnosis can often improve the survival rate of patients and simplifying the detection process is extremely important.In the traditional situation,the medical data of patients often requires manual detection and analysis by professional physicians.However,the proportion of doctors and patients in China is seriously unbalanced and the large amount of data is extremely demanding for the level and patience of professional physicians.In the process of detection and analysis,there may be a series of external factors,such as the status of testers,irregular data,randomness,and complexity of data,leading to uneven quality of detection and analysis which is disastrous for both patients and professional physicians.Therefore,seeking a method that can automatically search for biomarkers has become the appeal of many medical workers.With the rapid development of computer hardware and software,new machine learning and deep learning algorithms have made great progress in the fields of natural language recognition and image processing with remarkable achievements.Deep learning branched convolutional neural network has taken the lead in many fields by virtue of its powerful nonlinear representation capability.Inspired by the extensive application of convolutional neural network in other fields,this thesis proposed an automatic recognition algorithm of medical signals with high performance,strong robustness and high recognition rate based on the characteristics of each biomedical signal and focusing on the processing and analysis technology of biomedical signals.In this study,two kinds of medical signals,namely ECG signal and breast cancer pathological tissue image,were taken as the research objects for in-depth research.The main research contents are as follows:1.In the ECG signal recognition task based on the improved residual network,five kinds of most common heart disease data were selected from the MIT-BIH arrhythmia database.The waveform was positioned through the R-peak in the ECG signal waveform and the waveform was segmfied by 1×600 as a unit.Then,the segmented data was imported into the first convolution layer and the stacked residual blocks were used for feature extraction.Finally,softmax classifier was used for multi-case recognition.In the experiment,K-fold cross validation was used to train,verify,and test the MIT-BIH arrhythmia dataset.The proposed model was designed without any additional artificial features and data enhancement assistance.The accuracy of 97.20%,sensitivity of 92.85%,specificity of 98.29%,accuracy of 93.16% and F1 score of 93.00% were obtained.2.In the classification task of breast cancer pathological tissue images,traditional networks are unable to fully express the complex features of breast cancer pathological tissue images.In this thesis,a deep network model was designed by deepening the depth of the network to obtain deeper characteristic information to solve the problem of network degradation caused by the increase of model layers.The traditional convolution was replaced by dilated convolution which improved the perception field without bringing additional parameters.With the Adam optimizer,the learning rate decays exponentially with the training.In the experiment,dataset was from the Break His and a set of preprocessing frameworks was designed to improve the compatibility of the model.K-fold cross validation was used for the evaluation and an early stopping algorithm was used to avoid over/under fitting.An independent experiment on the dataset amplification factory was conducted and the proposed model was compared horizontally with existing models and literature.The experimental results show that the proposed model was efficient,accurate and reliable for the recognition of benign and malignant breast cancer pathological tissue images and for the fine identification of multiple cases in breast cancer pathological tissue images.