Lung four dimensional computed tomography (4D-CT) is of great value in tumor target localization and precise cancer radiotherapy. However, it is hard to segment tumors in 4D-CT data manually, since the data may contain a great number of slices with tumor. Meanwhile, auto-segmentation does not certainly guarantee the accuracy due to the complexity of images. Therefore, a new automatic segmentation technique based on Graph Cuts with star shape prior was proposed to increase automation and guarantee the accuracy of segmentation in our laboratory. Firstly, an object seed was selected in the image of initial phase and an initial target block was formed centering the selected seed. Then, the full search block-matching algorithm was adopted to obtain the most similar target block in the next phase and compute the motion field between them, and so on. Afterwards, the center seeds of each phase were obtained according to the motion fields, which would be set to the center point of star shape prior. Finally, tumors could be automatically segmented with Graph Cuts algorithm and star shape prior. Both qualitative and quantitative evaluation results showed that our approach could not only guarantee the accuracy of segmentation but also increase automation, compared with the traditional Graph Cuts algorithm.
Lung four dimensional computed tomography (4D-CT) can lead to accurate radiotherapy. However, for the safety of patients, the scan spacing of 4D-CT cannot be too small so that the inter-slice resolution of lung 4D-CT is low, and thus the coronal and sagittal images need to be interpolated to obtain high-resolution images. This paper presents a super-resolution reconstruction technique based on multi-model Gaussian process regression. We use the high-resolution transversal images and the corresponding low-resolution images as the training sets. The high-resolution pixels of the coronal and sagittal images can be predicted by constructing multiple Gaussian process regression models. The experimental results show that our method is superior to bicubic algorithm, projections onto convex sets, sparse coding, multi-phase similarity based method and Gaussian process regression method based on self-learning block in terms of the edge and detail recovery. The results demonstrate that the proposed method can effectively improve the quality of lung 4D-CT images, and potentially be applied to better image-guided radiation therapy of lung cancer.