Thorax Disease Identification Based On Deep Learning

X-ray is one of the most commonly used methods for diagnosing lung diseases.Compared with CT,X-ray is cheaper and has less radiation.In hospitals,hundreds of chest radiographs are produced every day.However,assessing a large quantity of chest radiographs is time-consuming and laborious for the radiologists.Therefore,develop-ing an effective automatic analysis algorithm for chest radiographs is of high demand.Thanks to the rapid development of computer technique and related image processing technology in recent years,computer-aided diagnosis has emerged and became a pow-erful tool for assisting doctors in the diagnosis of diseases.In order to improve the performance of the computer-aided diagnostic system in the screening of pulmonary diseases,this thesis has conducted some researches on chest X-ray images based on deep learning framework.The main research contents of this thesis are as follows:(1)A neural network with a priori perceptual mechanism is proposed to segment the lung field in this thesis.We first analyze the shortcomings of the existing U-Net.After that,in order to enable the network to effectively integrate the prior knowledge,we propose two schemes to generate the prior map to assist the model in segmenting the chest image.Besides,we also propose to use a spatial domain transformation net-work to correct the prior map in accommodating different samples,which makes the generated prior map insensitive to sample changes.We evaluated our algorithm on the JSRT(Japanese Society of Radiological Technology)dataset.The experimental results show that compared with the basic network,the proposed algorithm can effectively im-prove the segmentation performance of the model.(2)A relatively distance-sensitive neural network is proposed to classify multiple lung diseases.In traditional convolutional neural networks,the distance information is gradually lost as the networks become deeper.In order to effectively utilize the position information in the chest radiograph,we propose to make use of the relative position information to help the network diagnose diseases.In the proposed model,we first use the segmentation network to segment the chest-related tissue masks and extract the relative distances.The extracted information is encoded by the hourglass network,and the encoded distance features are merged into the main branch by the means of residuals.Compared with other distance information extraction and fusion strategies,the proposed scheme does not require manual intervention and has scale invariance.We verify our algorithm on the public ChestX-ray 14 dataset.The experimental results show that the detection performance of the trunk network for various pulmonary diseases can be effectively improved by adding relative distance features.Secondly,we also discuss the impact of network segmentation and fusion strategy on network performance according to our proposed model.