Development Of Computational Human Phantoms And Applications To Automated CT Image Segmentation

CT image segmentation is one of the most important procedures of tumor diag-nosis and treatment planning.The purpose of image segmentation is to divide the image data into organs and tissues.CT image segmentation can be achieved by time-consuming manual operations by doctors or physicists using a medical software system.Due to human introduced operating errors,treatment plan may suffer from adverse ef-fects of manual volume delineation.In comparison,automated image segmentation takes advantage of computerized analysis of pixel information or prior-knowledge,ca-pable of saving time and generating consistent contours.Model-based automated image segmentation technology uses prior-knowledge of the structure shape and deformable meshes.There are three main tasks for employing the method:structure model es-tablishment,registration between initial models and CT images,and optimization of mesh models.Computational phantoms-mathematical models of the human body-are frequently used in human anatomy simulations.since the 1960s,three generations of computational phantoms have been reported:(1)stylized phantoms(2)voxel phantoms and(3)boundary representation(BREP)phantoms.It is extremely inefficient to adjust the anatomical information in the voxel format,so the BREP phantoms-in the form of either non-uniform rational B-spline(NURBS)or polygonal meshes-became the lat-est research tool.Pregnant phantoms,obese phantoms,deformable phantoms and work phantoms were recently reported by using the boundary representation technology.The third generation phantoms contain large amount of deformable inner organs and tissues and are reasonable to be the initial guesses of model-based automated image segmenta-tion method.However,there is little literature on the use of phantoms in medical image segmentation.This thesis reports the development of 3D computation human phantoms for Chi-nese population and then the applications to automated image segmentation.To achieve the ambition,four tasks were divided:(1)development of a set of mesh-based age-dependent Chinese human phantoms,(2)extension of phantom features for CT seg-mentation tasks,(3)establishment of phantom based automated image segmentation procedure,and(4)development of releated softwares and integration with treatment planning system.We evaluated the initial input data.The missed data were completed using fitting methods.The original RPI-AM and RPI-AF phantoms,which represent adults,were scaled to yield new phantoms.Manual operations on a CAD software and some automated scripts were developed for adjusting organ or tissue meshes.A series of phantoms with different ages and genders were finished and named:USTC-AM,USTC-AF,USTC-15M,USTC-15F,USTC-10M,USTC-10F,USTC-5M,USTC-5F.Learning abilities with new images and contours,voxle features and CT features were added to phantoms due to the defferences between phantoms and CT images.Image segmenta-tions were finished by using CT features of phantoms and the active contour method.A series of adult CT images were used to test the proposed method.Segmentation re-sults of the brain,brain stem and spinal cord structures were found to match well with the ground truth.DSC values were higher than 0.7 and mean distance were around 2mm which are considered to be suitable for clinical cases.The voxel editing software,CT contours and labeled images conversion software,image registration software and automated image segmentation software were developed.In comparison with the man-ual method,our software reduced the time spent from 1 hour to 30 minutes,spending most time with the task of image registration.The registration program was success-fully imbedded into the DeepPlan-a commercial software.The results from the au-tomated image segmentation software were successfully imported into the 3D Slicer software suggesting that our results are compatible with clinically software tools.The main innovation of this work includes:(1)A more comprehensive set of anatomical data for the Chinese population involving different ages and both genders beyond those reported in current Chinese standards,leading to the development of a set of first-ever mesh-based age-dependent Chinese family phantoms,(2)Based on the previous set of phantoms,three additional features(automatic learning,voxelization,patient-specific CT image characteristics)were added,making it possible to propose and demonstrate a new phantom-based automatic image segmentation method which is successfully em-bedded into commercial tools to yield reasonable clinical results.