Research On Deep Learning Method For Automatic Detection And Classification Of Thoracic And Lung Diseases

Doctors can accurately and quickly detect and classify chest and lung diseases by using computers,which is of great significance for patients’ early clinical diagnosis and later treatment.In recent years,with the rapid development of deep learning technology,deep convolutional neural networks have been able to learn the semantic information and key features of diseases through training,and realize the analysis and processing of medical images.We mainly study the detection and classification algorithm of thoracic and lung diseases based on deep learning and its application.The main contents include:(1)We constructed an accurate dataset.Since the classification labels of the current largescale thoracic and lung disease datasets are basically analyzed from the diagnosis reports through natural language processing technology,the disease labels in the training set are only about 90% accurate.We construct a dataset of chest X-ray images based on the Chest X-ray 14 dataset and the NLM-CXR dataset,and configured accurate disease labels for each image.(2)We proposed a deep learning-based dual attention algorithm for disease diagnosis and classification of chest X-ray images.We utilize the feature channel attention module and the feature space attention module to selectively enhance the features that are highly correlated with pathology,while suppressing the useless feature information,thereby improving the feature extraction ability of convolutional neural networks.We use the self-built disease dataset and the Chest X-ray 14 dataset to verify the effectiveness of the proposed algorithm.Experiments show that the algorithm can achieve high-quality diagnostic results on both datasets.(3)We proposed a multi-scale feature fusion algorithm based on frequency attention network for analyzing chest X-ray images and diagnosing various thoracic and lung diseases.We use two-dimensional discrete cosine transform to extract and fuse the features of chest Xray images at different frequencies to reduce the information loss caused by down-sampling.In addition,we also designed a multi-scale channel attention module,which can capture multiscale cross-channel information interaction and enhance the performance and learning efficiency of channel attention.(4)We designed a deep learning-based auxiliary diagnosis system for thoracic and lung diseases.The auxiliary diagnosis system can accurately and quickly detect and classify the diseases that may be contained in the input medical images,and identify the areas where the diseases may exist,thereby assisting doctors in clinical diagnosis.We use the two algorithms proposed in this paper as the inference engine of the system.In order to improve the inference speed of the convolutional neural network,the algorithm model is processed by various optimization schemes such as layer fusion and model quantization.Starting from the basic characteristics of chest X-ray images and the principles of clinical diagnosis of related diseases,we propose two deep learning-based algorithms for diagnosing and classifying thoracic and lung diseases,and constructs an accurate labelled cardiomegaly disease dataset to complementary existing chest X-ray images datasets.At the same time,combined with the proposed algorithm,we designed an auxiliary diagnosis system for thoracic and lung diseases.The algorithm model of the system can basically meet the needs of clinical auxiliary diagnosis after relevant optimization.