Rib Suppression Of Chest X-ray Image Method Based On Generative Adversarial Networks

Chest X-ray is a commonly used imaging method for screening pulmonary diseases in clinic.However,the o verlapping rib structures in chest X-ray make it difficult to observe lesions that happen to fall in these areas.Suppressing the rib structure in chest X-ray and obtaining the soft tissue image obscured by the rib structure can improve the detection accuracy of doctors and computer-aided systems for lung diseases to a certain extent,which has important clinical significance.Dual energy subtraction is a common X-ray bone suppression technique in clinic at present,but there are many limitations in clinical practice because of the dependence on special equipment,high price and unstable image quality,and extra X-ray radiation.In view of the above problems,this paper mainly focuses on the rib suppression method of chest X-ray based on deep learning,The main work is as follows:(1)In order to solve the problem of inaccurate localization and recognition caused by overlapping and d complex rib structure and similar texture of different tissue structures in the existing rib suppression methods of chest X-ray,We propose a combined attention-based method for the suppression o f rib images on chest X-ray with generative adversarial network.Combining combined attentional mechanisms to obtain attentional maps of rib tissue in X-rays by generating adversarial networks.On the basis of retaining the lung field information in the original non-rib region,combining the rib tissue rattention map,adjusting the grayscale contrast between rib images and other human tissue structures in the chest,and acquiring rib suppressed X-ray chest images.Instead of generating bone suppression images directly,we reduce the presence of ribs by generating rib attention maps and fusing the original image with the attention map to achieve the purpose of suppressing rib images in chest X-ray.We not only avoid generating rib images directly,which may lead to incomplete rib images and poor bone suppression results,but also avoid generating soft tissue images directly,which may lead to blurred images and loss of details.The performance of the method in this chapter was evaluated by qualitative and quantitative experiments on the public X-rays chest datasets—JSRT.The experimental results showed that,compared with the SOTA method,although the method in this chapter had little difference in the index of SSIM,the index of PSNR was improved by 5 points,and the comprehensive comparison of our method is significantly better than the comparison method.(2)In order to solve the problem of limited feature information extraction capability in the above-mentioned X-chest rib image suppression method,resulting in insufficient amount of generated bone suppression feature information,We propose a joint predictive filtering based generative adversarial network for rib suppression in X-rays.On the basis of the above proposed combined attention-based X-ray rib suppression network,the depth predictive filtering technology is introduced to to perform feature-level filtering from after the encoder of the generation network and image-level filtering from after the decoder,respectively,to enhance the contextual correlation between the information,add more detailed information,and further enhance the network performance to achieve better results on the processed information.At the same time,we added the feature matching loss function to further constrain the bone suppression image generated by the network from the feature level,and further accelerate the speed and precision of the convergence of the network towards the target of rib suppression in X chest radiograph.The experimental results show that our method has better rib suppression effect compared with the existing mainstream methods.The method in this chapter achieved the highest values for both PSNR and SSIM metrics,42.661 and 0.990,respectively.