Registration Based Super-Resolution Reconstruction For Lung 4D-CT Images

Four-dimensional computer tomography (4D-CT) is becoming widely used in lung cancer treatment for providing respiratory-related information that is essential for guiding radiation therapy. Lung 4D-CT data is usually obtained by simultaneously collecting CT segments and respiratory signals, and then sorting the multiple CT segments according to different phases and reconstructing three-dimensional images. Lung 4D-CT can represent not only the shape of organs and targets in thorax during the respiratory cycle, but also reflecting the extents and features of the movement. It has a profound effect in precise radiotherapy for moving targets. The lung 4D-CT images show the locations of targets and organs when breathing, and help to estimate their movements, which is great helpful for designing the individualized treatment plan with the patient. This individualized plan can reduce the irradiated area and increase the target dose, and at the same time spare the normal tissues. However, when collecting 4D-CT data, the more scanning time, more irradiation dose absorbed with patients. Patients need repeatedly prolonged scanning duration in a couch (at least one respiratory cycle), and the radiation dose required by 4D-CT has an order of magnitude higher than the dose required by 3D-CT once a time. In order to avoid excessive dose, inter-slice spacing is widened to reduce the number of scanning slices, resulting in 4D-CT data acquired with high inter-slice spacing and an inter-slice thickness that is much greater than in-plane voxel resolutions. Therefore, when observing the right proportion multi-planar images, the simple and effective operation is implementing interpolation along superior-inferior direction according to the proportion of inter-slice resolution and in-plane resolution. The conventional interpolation methods are the nearest interpolation, bilinear interpolation, cubic interpolation and spline interpolation. However, these simple approaches do not introduce new information, and can not recover the high frequency information in the high-resolution image, which would result in blurring image. To improve the quality of the multi-planar display images, in this paper, we study the registration based super-resolution technology to reconstruct high resolution lung 4D-CT images.Image super-resolution reconstruction technique is based on giving a single low-resolution image or multi-frame low-resolution observed images corresponding to the same scene which has sub-pixel movements between the images, and then integrate the redundancy information in the single image or non-redundant information between multiple images to reconstruct the high-resolution image. This is an effective post-processing algorithm to improve image resolution. Lung 4D-CT data provides low-resolution images of different phases with movements between them, and each phase corresponds to different movement time of the lung. Therefore, the multi-planar images of the different phases can be considered as "frame" about the same scene. The "frames" have the similar structure, and there are sub-pixel displacements between them. So the characteristics of lung 4D-CT data meet the demand of super-resolution reconstruction technique, and this paper adopts the super-resolution reconstruction technique to reconstruct high-resolution images, which is the basic idea of this paper.Based on the above idea, this paper proposed a registration based super-resolution reconstruction to improve the resolutions of lung 4D-CT multi-planar display images. Most of the image super-resolution reconstruction techniques are based on the image degradation model. Firstly, an initial high-resolution image is estimated. Then estimating the motions between the low-resolution images and the point spread function (PSF) related to image blurring, down-sampling and movements. Finally, iteratively recover the estimated initial image to get the super-resolution image. The low-resolution observed images provided by 4D-CT data are the degradation images. In this paper, we analyze the basic model of image degradation first. Then, we employ Active Demon registration method to estimate motion vector field of different "frames". Finally, based on the motion vector field, we employ the projections onto convex sets (POST) algorithm to reconstruct lung high-resolution images. At the last, further research is done to improve the POCS algorithm:optimizing the high-resolution initial image via NEDI method and improving the PSF to reduce edge artifacts. The main work of this paper is as follows:(1) Description of the image degradation model. In realistic imaging process, the rich information in the original scene can not be fully revealed, which results in a group of low-resolution observation images obtained. There are many factors affecting the final image quality, such as deformation caused by the relative motions between the imaging system and the original scene, image blurring caused by the relative motions and the atmospheric turbulence, inadequate sampling limited by the hardware, and the unavoidable noise. Therefore, the observed low-resolution images can be taken as a degradation version of the original high-resolution images obtained through deformation, down-sampling, sensor blur and noise. This process can be simply mathematical and modeled associating the high-resolution image and the low-resolution observed images. The image super-resolution reconstruction is the inverse problem of the degradation process, to reconstruct a high-resolution image through multiple low-resolution images.(2) The use of Active Demons registration algorithm to obtain motion information between different phases. For lung 4D-CT data, the movement of lung is limited, so the elastic registration algorithm can be employed to obtain motion information. Motion estimation is a key step in the SR process. Motion estimation with sub-pixel accuracy is required in the reconstruction process. This paper uses Active Demons registration algorithm to obtain motion information, which is based on the optical flow field, and applying the classical Maxwell’s demons thermodynamic principles to the image registration. This elastic registration method is sensitive to the small deformation following the time-varying. The method is relatively simple, high precise and high speed, without any pre-processing of the image sequence. In this paper, we firstly employ Active Demons registration algorithm to estimate the motion fields between different phases, and the results show this method can accurately estimate the lung movement.(3) Based on the motion vector field, we employ the projection on convex set (POCS) algorithm to reconstruct high-resolution lung images. Super-resolution method can be roughly divided into two categories:frequency-domain method and spatial method. Frequency domain method is based on the Fourier transform, removing the spectrum aliasing of low-resolution images converted to the frequency domain. However, this method is difficult to deal with the complexity model with serious deformation and blur, and is strict with noise which is not suitable for lung 4D-CT data. The spatial method is more suitable for the actual application since it considers global and local movements, spatially variant blur, sensor noise, arbitrary sampling and so on. So we select POCS method that is a spatial based super-resolution reconstruction formulation to reconstruct high-resolution images, because it is simple, easily added a priori constraints, and can retain image edges and details better. The results show clear reconstructed lung images with improved structure and details.(4) The reconstruction results via POCS algorithm have ringing artifacts around edges, so further improved algorithm is proposed. We analyze the reasons of edge artifacts during POCS approach. Since the point spread function (PSF) is usually based on isotropic Gaussian function, which is not accurate when the pixels in edge region is estimated. Due to isotropic Gaussian function, the pixels with high value are corrected with higher value, and pixels with low value are corrected with lower ones, resulting in the ringing edge. So NEDI is used to generate an initial estimate of high-resolution image with the clearer and sharp edge instead of the cubic spline interpolation method, and anisotropic based Gaussian function which relate to PSF in the degradation model is employed to reduce the edge ringing and improve edge detail quality. The improved algorithm has profound results with improved edge and reduced edge ringing.The data used in this paper are obtained from a publicly available dataset that was provided by the DIR-lab at the University of Texas M.D. Anderson Cancer Center. The dataset consists of 10 groups of lung 4D-CT data, each group contains 10 phases. We evaluate our method based on simulated data and real data respectively. The dataset only provide high-resolution transverse images. So we simulate low-resolution images using high-resolution transverse images for super-resolution reconstruction and evaluate the proposed method using RMSE. Meanwhile, the low-resolution coronal and sagittal images are used to reconstruct high-resolution images and evaluation the proposed method using edge width and average gradient value. The experimental results show that the proposed algorithm profound a clear high-resolution image, significantly decreases RMSE values and edge widths, and increases image average gradient values compared with the conventional methods such as cubic spline interpolation and BP algorithm. The further improved algorithm outperforms the original one. The reconstruction results show that both edge artifact and RMSE values decrease. All the results demonstrate that the super-resolution reconstruction method proposed in this paper can effectively improve the resolution of lung 4D-CT multi-planar display images.