Low-rank Matrix Reconstruction And Multi-model Gaussian Process Regression Based Super-Resolution Reconstruction For Lung 4D-CT Images

Respiratory motion can lead to the movement of organs and targets in thorax,which takes some difficulty in individualized precise radiotherapy.In order to providing respiratory-related information for high-precision radiation therapy,lung four dimensional Computed Tomography(4D-CT)has been playing a more and more important role in recent years.Lung 4D-CT images not only truly reflect of the respiratory cycle in the shape of the thorax organs and tumors,but also reflecting the extents and feature of lung movements.Doctors can well determine the shape of tumor and the range of motion due to the lung 4D-CT images.However,during the 4D-CT scan of the lungs,the patient needs to make prolonged scanning repeatedly in the same couch,and dense sampling would be unfeasible among the superior-inferior direction in case of excessive radiation dose of the patient.Therefore,inter-slice spacing is widened and the number of scanning slices is reduced during the lung 4D-CT scanning,and thus the inter-slice resolution of the image is much lower than the in-plane resolution,which results in anisotropy of the data.In order to display the right proportion of the multi-planar images,it is necessary to interpolate along the superior-inferior direction.However,the interpolation does not introduce new information,and thus the image tends to be blurred easily.In order to solve this problem,this paper presents two methods based on super-resolution technology to reconstruct high-resolution images.The main works are as follows:1.We propose a low-rank matrix construction based super-resolution reconstruction algorithm:For one low-resolution image along superior-inferior direction in a position,other phases of the images in the same position can be considered as the multiple "frames"which have similar structure and sub-pixel displacements.The degradation process of high-resolution images to low-resolution images can be seen as the process of the destruction or contamination of certain elements of the matrix.Because the positions of destruction or contamination in different "frames" are not the same,we can use information in other "frames" to complete the observed low-resolution image.Based on the above idea,we sample the patches in raster-scan order in the observed image and find the most similar patches of the corresponding image patches of the other "frames"using the similarity measure function.Then we construct an approximately low-rank matrix and recover its low-rank character by iterative soft thresholding algorithm.Then re-integration to obtain the reconstructed patches of the target image.We consider that the reconstructed patches can exploit the information of patches from other phases and can reduce the artifact absence of other phases to increase the resolution.Through the processing of all the patches from the observed low-resolution image,we can obtain the complete high-resolution image.2.We propose an algorithm based on multi-model Gaussian process regression to achieve super-resolution reconstruction:The method of low-rank matrix construction needs to find similar patches,and the coronal and sagittal images utilized in the method are low-resolution images.The above reasons limit the speed and accuracy in the process of algorithm.Current learning-based approaches can be implemented by training one or more complete models from the training set to obtain high-frequency information.Introducing prior knowledge from the high-and low-resolution images in the training set can help recover some details in the observed low-resolution image and increase the resolution.Based on the above ideas,we utilize high-resolution transverse images and corresponding low-resolution interpolation images as the training set,divide the training set into multiple groups,and introduce Gaussian process regression(GPR)to train multiple GPR models.Therefore,utilizing the observed low-resolution coronal or sagittal images as well as GPR models trained from training set,we can reconstruct high-resolution images.In this paper,two publicly available lung 4D-CT data sets are used to verify the proposed methods.In the simulated experiment,the low-resolution images we used to reconstruct are degraded from the high-resolution transverse images.Meanwhile,we evaluate the reconstruction results from the visual effect as well as quantitative evaluation,and carry out the correlation parameters analysis by simulated images.Experimental results illustrate that two proposed methods can reconstruct more precise high-resolution images compared to other methods.In real images experiment,we deal with some low-resolution coronal and sagittal images using this two proposed methods and evaluate the reconstructed high-resolution images visually.Experimental results show that our methods can produce clear edges as well as details and improve the image quality.Meanwhile,the proposed methods significantly decreases edge widths compared with the conventional methods.