Bone Suppression Methods Of Chest Radiographs Based On Deep Convolutional Networks

Chest X-ray is a common diagnostic screening tool.However,due to the overlapping anatomical structure,the common chest X-ray has caused many obstacles on clinical diagnosis and makes it more difficult for computer-aided analysis and diagnosis.To reduce the diagnostic barriers caused by bone structures,bone suppression has been achived in chest X-rays,which can provide help for manual or computer-aided diagnosis of lung diseases and reduce the rate of misdiagnosis.The research goal of this paper:is to conduct the bone suppression of a single chest X-rays by self-developed computer algorithm.We use the dual-energy subtraction data,combined with the characteristics of bone component,the convolution neural network has been t.rained and a bone prediction model which could generate soft tissue images is constructed.The soft tissue images could be provided for clinician diagnosis as a reference.The main contents of the thesis are as follows:(1)To achieve bone suppression of chest X-ray,we propos cascade of multi-scale convolutional network(CamsNet).We use deep convolutional neural networks(ConvNet)as the basic prediction unit for the mapping between the gradient images of chest X-ray and the corresponding gradient images of bone.The CamsNet in gradient domain is constructed by multiple ConvNets in a cascaded structure.With the resolution increased,the gradient images of bone are gradually predicted by different ConvNets in different scales,and the final bone image is fused from gradient images of bone in various scales by the framework of maximum-aaposteriori probability.The estimated bone image was subtracted from original chest X-ray to obtain the final soft tissue image which is also a bone suppression images.This method produces high quality and high resolution bone images and soft tissue images.We can apply X-ray chest images collected by different devices to the trained model,and also get visually clear bone and soft tissue images.The CamsNet model can be used on the chest X-rays acquired by various types of X-ray devices,for example,scanned films,and can also produce visually appealing bone and soft-tissue images.(2)Cascaded convolutional network model in wavelet domain(Wavelet-CCN)is used to construct a cascaded model to conduct bone suppression of chest X-ray.The sparsity of the wavelet coefficients is suitable as an output of convolutional network.The convolutional network uses the wavelet coefficients of chest X-ray to predict the wavelet coefficients of the bone image.By combining multi-level wavelet decomposition and cascaded structure,the Wavelet-CCN model can gradually predict bone in terms of precision and spatial resolution through multi-scale methods.Compared with the previous CamsNet model in gradient domain,the Wavelet-CCN model predicts the wavelet coefficients of the bone first,and then reconstructs the bone image by inverse discrete wavelet transform.This method can avoid the insistent background caused by 2D integration of gradient images.(3)We observe the effect of bone suppression on lung nodules in the original chest X-ray and computer-aided diagnosis of tuberculosis.The degrees of contrast between the nodules and the background,and the intensity of the nodules are quantitatively evaluated before and after the bone suppression of chest X-ray,Also the computer-assisted diagnosis of tuberculosis improved by bone-suppressed chest X-ray is also evaluated by ChexNet.In this paper,the bone suppression algorithm is systematically studied.Two convolutional network models in gradient and wavelet domains are proposed.And using the bone suppression of chest X-ray,a significant improvement in computer-aided diagnosis of tuberculosis is observed.