X-ray Image Super-resolution Based On Multi-feature Learning

High-resolution medical X-rays provide more lesion detail,which is essential for medical diagnosis.For example,lateral head x-rays are used to diagnose bony or dental malocclusion.However,the quality of Radiograph images is limited by scanning time,patient posture,movement,etc.,and obtaining higher-resolution Radiograph images is bound to change hardware conditions and is expensive.Therefore,it is of great clinical significance to improve the quality of Radiograph images by using image super-resolution reconstruction technology.Existing methods lack prior information related to anatomy,and it is difficult to accurately capture the texture details of X-ray images,and super-resolution reconstruction of X-ray films is more challenging than natural scenes.In view of this problem,the main innovations and contributions of this paper are as follows:1.Aiming at the complementary information between the similarity of medical anatomical structures and the super-resolution of images,a geometric iterative collaborative network method is proposed.The network realizes the super-resolution reconstruction and key point detection tasks of X-ray image at the same time,extracts the feature information of the key point detection network and the super-resolution network and performs feature fusion:the key point detection network extracts the global heat map features of the key point;The super-resolution network uses the generative adversarial network to fuse the key point heat map predicted by the network to generate clearer high-resolution X-ray images.Experiments verify the effectiveness of geometric iterative collaborative networks on Cephalometric and Hand datasets.2.For medical X-ray images with pose differences,a contextual local consistency constraint method is proposed.In this method,a generative adversarial network learning image based on the Axial Transformer framework is designed.The generative adversarial network is used to reconstruct high-quality X-ray images,while fully considering the overlapping image characteristics of X-ray images,a local perception constraint module is proposed,and the residual artifacts between the reconstructed image and the original high-resolution image are statistically calculated to constrain the network to generate real texture details,suppress the generation of artifacts,and further ensure the local consistency of the image.Experiments demonstrate the effectiveness of the method on the Mura and Chest datasets.3.According to the two proposed super-resolution reconstruction methods,a super-resolution reconstruction system suitable for medical images is designed and implemented.The system is simple to use and can be used for X-ray image reconstruction and positioning of key points on the head and hands.Users can reconstruct high-resolution images from low-resolution medical images to better present details and structural information.In the reconstruction of X-ray images and the positioning of key points of the head and hands,the super-resolution system proposed in this paper has excellent performance,which can not only assist doctors in diagnosing X-ray images,but also provide strong support for the research and application of medical imaging.