Texture Analysis And Delineations Of Head And Neck Cancer Based On PET/MRI Images

Currently, head and neck cancer(HNC) has been the world’s sixth-largest high incidence of cancer and a serious threat to people’s health. Precision conformal radiation therapy techniques such as intensity modulated radiation therapy(IMRT),which tightly conforms the radiation dose to a target, can potentially improve local-regional tumor control, reduce normal tissue toxicity, and improve the quality of life for the patients with HNC. The key to the successful implementation of high-precision radiotherapy is the precise delineation of tumor target volume based on positron emission tomography(PET), computed tomography(CT), or magnetic resonance imaging(MRI) images. Its essence is the segmentation of PET/CT/MRI images. In this paper, according to the features of PET, MRI images of the HNC and target segmentation problem, we have made the following research on texture analysis and segmentation of tumors.(1)The standard uptake value(SUV) of HNC is usually very close to that of its surrounding normal tissues, which makes it difficult to separate the tumor tissue and its adjacent normal tissue relies solely on the information of PET SUV. We found that the contrast texture feature of the PET images based on neighborhood gray-tone difference matrix(NGTDM) is helpful to solve this problem, and proposed an improved random walk(RW) algorithm for PET image segmentation of head and neck tumor based on PET SUV and contrast texture feature of tumors. Firstly, the selected region of interest(ROI) was segmented into the primary tumor, normal tissue and pending regions by three-dimensional adaptive region growing and morphological dilation on PET SUV images. At the same time, the primary tumor and normal tissue regions were labeled as foreground seeds and background seeds for RW, respectively.That not only made RW more efficiently, but also made segmentation results more accurate. Secondly, due to different contrast texture features of head and neck tumor and surrounding normal tissue in PET images, the contrast texture futures were incorporated into the weights of RW to further improve the accuracy of tumor segmentation results.(2)PET images can provide the molecular biology functional information such as metabolism, proliferation and hypoxia in tumors. MRI images can provide anatomic structure information. We found that combine functional information from PETimages and structural information from MRI images always lead to a better segmentation result. Therefore, a two-stage adaptive region growing(ARG) algorithm was proposed to automatically delineate tumor target volumes(TTV) for PET and MRI-guided radiotherapy treatment planning of head and neck tumors. Firstly, the voxel with the maximum SUV value in the tumor volume of interest was selected as the initial seed of ARG. The first stage ARG was applied to PET images and MRI images, respectively. Secondly, the minimum SUV value in the resulting TTV based on PET images and the best threshold for the first stage MRI ARG were combined to determine the growth criterion of the second stage ARG based on both PET and MRI images. The proposed method can effectively differentiate tumors from edemas and normal tissues.(3)We extracted 11 kinds of texture features based on the gray level run-length matrix. Through a lot of experiments, we found that the long run emphasis(LRE)texture and long run high gray-level emphasis(LRHGE) texture of the MRI image was useful to identify the tumor region.