Tuberculosis Detection And Classification From Pulmonary X-rays Based On Deep Convolutional Neural Network

Tuberculosis is the world's deadliest infectious disease,but it is curable.The main challenge is early detection and classification of tuberculosis,which is a critical step in the treatment of tuberculosis.The existing computer aided detection systems have had preliminary researches on diagnosis of pulmonary tuberculosis for chest X-rays,but there is still a lack of a more in-depth analysis,such as the location of tuberculosis lesions area,and the classification of focal signs.With the development of deep neural networks in recent years,using deep convolutional neural networks to solve the above problems has great advantages and potential.In this thesis,several improved deep convolutional neural networks for the classification and detection of pulmonary tuberculosis are proposed for chest X-rays.The main contents are as follows:(1)A tuberculosis data set containing 5299 chest x-rays is builded,including 2,383 images without disease and 2,806 images with tuberculosis infection.The chest X-rays of pulmonary tuberculosis infection are marked by 3 professional radiologists,including 6 types of pulmonary tuberculosis signs.(2)A method of lung parenchymal segmentation based on U-Net is proposed.Aiming at solving the problems that there are few data for lung parenchyma segmentation on X-rays,the data augmentation operation is adopted for the segmentation data at first.Then we uses U-Net to complete the lung parenchymal segmentation on from X-rays according to the characteristics that U-Net is very effective in the training of small sample data.The experimental results show that the Dice and the Jaccard are 0.971 and 0.944 respectively,which exceed those of frontier algorithms in this field.(3)We proposed an improved Faster R-CNN network for focus detection,which solves the problem that the differences of pulmonary tuberculosis lesion areas are too large and lead to dissatisfied results of lesion detection.We add a feature pyramid structure in Faster R-CNN and introduce nonlinear fusion operation of feature maps in the network.The feature map of each layer of the network is guaranteed to have rich semantic information,and the detail information of the X-rays is hardly lost.The proposed method improves mAP(mean Average Precision)by 62.4% compared with Faster R-CNN.(4)A classification algorithm of chest X-rays based on the detection results of tuberculosis lesions is proposed.Compared with the Acc and AUC of the classification method based on Dense Net proposed in chapter 3,experimental results on the TCLD_CXR data set show that the Acc and AUC of this method are improved by 10.6% and 11.6% respectively.In the conclusion,we improve deep convolutional neural networks and achieve the tuberculosis detection and classification from chest X-rays,which reduce the workload of doctors and provided good suggestions for the diagnosis of tuberculosis.