Based On Multi-atlas Deformable Registration To Auto-segmentation Organ-at-risks(OAR) For Naso- Pharyngeal Carcinoma(NPC) Radiotherapy

Along with the development of computer technology, radiotherapy also developed from 2-dimension(2D) routine treatment to 3-dimension(3D) precise treatment. In the 3D precise treatment, the accurate drawing of target and OARs were required. There are some problems using manual contouring, for example, time-consuming, large intra-and inter- domain expert variability and so on. In clinical, we need some methods to produce OARs quickly and precisely. NPC were high risk in southeast China, this tumor region was one of the most important and complex organs within human body. The OARs for NPC of automated segmentation were significant for radiotherapy in clinical.In this thesis, firstly, the internal and external situation of radiotherapy for NPC is introduced and discussed of the need of automated segmentation in clinical. Moreover, the basic line of investigation for automated segmentation is given. Secondly, three steps of automated segmentation for OARs in NPC is presented. During the first step, the details of basic linear registration and the relative method of evaluation for registration are given. It also gives the basic components of the registration framework, included transform, interpolator, metric and optimizer. At the same time, EMMA algorithm is presented. Linear registration is realized by ITK tools. During the second step, the mainly introduction is poly-smooth nonlinear registration algorithms, including Log-Euclidean poly-affine transform(LEPT) and block match. And how to use block match method to evaluate the result for poly-smooth nonlinear registration. The last step introduces shape-constrained dense deformable registration. It gives the details of local correlation coefficient(LCC).Then a full free-form deformable registration also realizes using ITK tools. At end of this thesis, STAPLE algorithm is introduced for fusion of multiple atlas segmentations. It presents the method of selected atlas and uses three methods to evaluate the result between auto-segmentation and manual contour, including DICE similarity coefficient index, distance transformation and difference volume calculation.In practice, selected 7 patients image that expert had manual contoured all OARs as atlas database in two different hospitals, and 15 new patients is auto-segmented by the application that programmed in this thesis. Comparing the auto-segmentation and expert manual contouring use the above three methods to evaluate. The test proved that overall contouring efficiency and accuracy can be significantly improved.