Key Techniques For Classification Of Thorax Diseases Based On Deep Learning

In the face of the explosive growth of medical imaging data and the serious shortage of radiologists,the search for fast and accurate intelligent diagnostic methods is a relentless effort of researchers in China and abroad.With the remarkable progress of deep learning technology in the fields of object detection,image classification and image segmentation in recent years,it has promoted the development of medical image processing fields such as disease classification,lesion detection,segmentation,registration and image annotation.This makes the chest X-ray computer-aided diagnosis(CAD)system one of the hot spots in medical imaging research,and has gradually entered medical clinical applications.Thoracic disease is one of the most common diseases in the world,with lung cancer being the third most deadly disease worldwide.In addition,abnormalities such as pneumothorax,effusion and infiltration are also early signs of some more serious lesions,and early detection and timely treatment of these abnormalities or lesions can prevent other more serious chest diseases and effectively reduce mortality.Therefore,it is of great practical significance to combine computer vision with medical clinical applications and to carry out research on auxiliary diagnosis systems for thoracic lesions.Chest X-ray is currently the best method for diagnosing chest diseases.Compared with CT examinations and ultrasound examinations,chest X-ray is cheaper,less radioactive,and faster and more convenient to operate,so it is usually used in the preliminary stage of chest disease examination and is most widely used to examine lung cancer,tuberculosis,pneumonia and other pathologies.Although modern hospitals collect and store a large number of X-ray images and diagnostic reports,building a deep learning model based on these big image data to build a high-precision computer-aided diagnosis system still faces the following challenges: First,the features of some thoracic diseases in X-ray images are often blurred and appear to be easily disturbed by thicker soft tissues resulting in enhanced opacity,which can confuse some more serious diseases with other benign abnormalities and cause misdiagnosis;the second is even the feature performance of the same disease on different patients’ radiographs varies greatly,which will increase the learning difficulty of the network;secondly,the fine-grained pixel-level annotations used for model training are insufficient,the global X-ray images have more redundant features,and some non-focal regions are not helpful for the diagnosis of professional radiologists;in addition,the existing machine learning algorithms cannot obtain the localization of focal regions,resulting in low clinical utility.In response to the above challenges,this paper proposes an AC-DSENet model that combines image registration technology,image segmentation technology and image classification technology,and uses it to classify 14 types of thoracic lesions.In order to solve the problem that the features of X-ray images are fuzzy and easily confused with benign abnormalities,a dense squeeze excitation network is proposed.By combining the squeeze excitation module with the dense connection block,the classification network can pay close attention to the different channels of the X-ray image feature map.Therefore,the expression of useful information in the network is enhanced and the transmission of useless information is suppressed,so as to enhance the classification performance of the network for chest diseases.In order to solve the problem of large differences in chest X-ray performance of thoracic lesions in different patients,a focal loss function is proposed to replace the binary cross-entropy loss function used in the traditional multi-label classification task to calculate the loss of each lesion category.By introducing hyperparameters,it will be difficult to identify.The lesions that are more likely to be multiplied by a larger weight,and the easily identifiable lesions are multiplied by a smaller weight,allowing the network to pay more attention to the learning of complex lesion types.In order to solve the problem of many redundant features of the global X-ray image,it is proposed to apply the image registration algorithm to the multi-atlas segmentation in the image preprocessing stage to realize the automatic segmentation of the classification data set,and only keep the heart and two parts of the chest X-ray.For the lesion area of the lung lobe,the segmented images are used to train the classification network to suppress the learning of redundant features and improve the training speed and efficiency of the classification network.In addition,the classification results of the disease are output in the form of an intuitive heat map,and the highlighted lesion area is to prompt the professional physician to focus on the observation of the area,so as to achieve the role of auxiliary diagnosis and increase the clinical practical value of the AC-DSENet model.Experiments were performed on the ChestX-ray14 dataset,and the results showed that the classification accuracy of AC-DSENet in 14 chest lesions was significantly improved compared with the existing three classical algorithms,and the average AUC value reached 0.858.And by adding segmentation constraints,the redundant features learned in the data training process are effectively reduced,and the training time of the classification network is shortened by 39.8%.At the same time,this paper also locates the lesion areas diagnosed by the model for visualization and display,which meets the practical needs of assisting doctors in diagnosis.