Super-resolution Reconstruction For Lung4D-CT Coronal And Sagittal Images Based On Motion Estimation

Precise radiation therapy based on lung4D-CT image is the current effective treatment for lung cancer because of its capability in providing a comprehensive characterization of respiratory motion for high-precision radiation therapy. With4D-CT, additional phased images can help capture respiratory motion information that is crucial for target definition in radiation therapy.Therefore lung4D-CT is playing an increasingly important role in lung tumors precise radiotherapy. Lung4D-CT is usually obtained by sorting the multiple free-breathing3D-CT segments in relation to the couch position and the tidal volume. However, due to the inherent high-dose exposure associated with CT, dense sampling along superior-inferior direction is often not practical, thus resulting in an inter-slice thickness that is much greater than in-plane voxel resolutions and anisotropic data.Therefore, when observing the coronal and sagittal images of each3D data, we must implement interpolate operation along superior-inferior direction according to the proportion of inter-slice resolution and in-plane resolution to obtain correct image display. The nearest interpolation and bilinear interpolation are conventional methods, however, these approaches will blur the image, especially when the proportion of inter-slice resolution and in-plane resolution is large.To solve this problem, this paper employs super-resolution technology to reconstruction clear lung4D-CT coronal and sagittal images. Super-resolution technique is the technology that can reconstruct a high-resolution image by use of multiple frames low-resolution degraded images in the same scene with mutual displacement. Lung4D-CT data is composed of10-20different phases of3D images, and each phase corresponds to different movement time of the lung.Therefore, for one low-resolution lung coronal(sagittal) image (along superior-inferior direction) in a position of a3D image, the other phase coronal(sagittal) images in the same position can be considered as the multiple "frame" low-resolution images of the same scene. Consequently, super-resolution techniques can be used to obtain high-resolution lung4D-CT coronal(sagittal) images. This is the basic idea of this article.Based on the above idea, we propose super-resolution reconstruction method based on motion estimation to improve the image resolution of lung4D-CT coronal and sagittal images.First, we analyze the basic model of image degradation.Then, we employ motion estimation method based on full search block matching motion estimation and fast sub-pixel motion estimation respectively to estimate images motion vector field of different "frames", respectively. Finally, based on the motion vector field, we employ the iterative back projection (IBP) algorithm to reconstruct high-resolution lung4D-CT coronal(sagittal) images.The main work of this paper is as follows:(1) Analyze the basic model of image degradation.In realistic imaging process, we obtained a set of low-resolution image due to the unavoidable deformation caused by geometric transformation such as translation and rotation movement, optical blur, inadequate sampling and introduced noise. Therefore, the observed low-resolution image is a degradation version of the original high-resolution image through a series of degradation processes.The super-resolution image reconstruction technique is the inverse problem of the degradation process, which means it can reconstruct a high resolution image by combining multiple low resolution images.(2) Apply motion estimation algorithm to obtain lung motion information. With4D-CT, composed of10-20different phases of3D images, and each phase corresponding to different movement time of the lung, additional phased images can help capture respiratory motion information that is crucial for target definition in radiation therapy. This paper uses block-matching motion estimation algorithm to obtain lung motion information, the algorithm principle is relatively simple, and not require any pre-processing of the image sequence, only requires calculate the original images, and thus retain all information of the image sequence. Our experiment employ motion estimation algorithm based on full search block matching to estimate image motion vector field between different phases of lung coronal and sagittal images, and lung movement trend can display accurately.(3) The problem of full search block matching motion estimation is slow operation, large amount of computation, and the search accuracy depends on the step length.Therefore, this paper studies a fast sub-pixel motion estimation algorithm to improve the speed and accuracy of motion estimation further. It combines the fast three-step search method and optical flow method. The results show that image not only motion vector is more accurate, but greatly reduce the computation time compare with full search algorithm before.(4) Based on the motion vector field, we employ the iterative back projection (IBP) algorithm to reconstruct high-resolution lung4D-CT coronal(sagittal) images. Super-resolution method can be roughly divided into two categories:frequency domain method and spatial method. However, frequency domain method is difficult to deal with the problem of image noise, and can only deal with the degradation models with global movement. Therefore, this method is not commonly used due to great limitation. So we select super-resolution reconstruction based spatial—iterative back projection (IBP) method, it have simple principle, flexible form, and the important point is better to retention effect of image edges and details.We use a public available dataset to evaluate the proposed super-resolution reconstruction algorithm, the dataset consists of10groups of lung4D-CT data, each group contains10phases. We select the coronal and sagittal images of different phases to perform above procedure for experiments. Experimental results show that the proposed algorithm significantly decreases edge widths compared with the conventional methods such as nearest interpolation and bilinear interpolation; significantly increases image average gradient value compared with bilinear interpolation. Meanwhile, this paper implements a comparison of reconstruction results based on full search block matching motion estimation and fast sub-pixel motion estimation, respectively. The results show that the latter have clearer reconstruction. More important, the computation time reduced nearly20times than the former.