Automatic Thoracic Disease Localization And Recognition Based On Neural Networks

Chest X-rays play a key role in the medical diagnosis of chest-related diseases.For many chest diseases,such as pneumonia,tuberculosis,and enlarged heart shadow,Chest X-ray is the preferred method for early screening and later review.However,the interpretation of chest X-rays is not easy and usually requires clinically experienced doctors.Therefore,it is of great value to develop systems with deep learning for automatic identification of chest X-rays,which can effectively assist doctors in completing the work and improve their work efficiency.At present,there are many algorithms that apply deep learning to the automatic recognition of chest X-rays.However,due to the particularity of chest X-rays as medical images and the lack of bounding box data sets,it is still challenging to achieve automatic chest diseases detection and localization.Based on this,we built a bounding box data set,named Chest-box,which contains 3952 positive CXR images and 9960 bounding boxes.Then,using the Chest-box data set as the training set,we proposed an RPN network embedded with FPN as a region proposal model to extract the region of interest(ROIs).Finally,taking the ROIs as attention information,we further proposed an attention mechanism network combining Dense Net and attention mechanism to achieve the disease detection and automatic localization of chest X-rays.The testing on the Chest X-ray14 data set shows that compared with previous work,our approach has the state-of-the-art performance in both classification and localization of diseases.The dissertation is organized as follows:Chapter one mainly introduces the background and meanings of our research.Chapter two is a brief review of theoretical knowledge related to chest X-rays and deep learning networks.Chapter three describes our deep learning network model and its results.Chapter four is a summary of this dissertation.