Research On Construction Of Deformable Statistical Head Atlas And Automated Brain Parcellation

Digital anatomical atlas of the head relies on modern medical imaging and computer modeling technology to digitize and visualize the anatomical structure information of the head.It provides accurate and complete digital anatomical models.Most of the existing digital anatomical atlases of the head are constructed based on a single and specific anatomical morphology of human,which have been widely used in medical research and anatomy education.However,in the fields of personalized diagnosis and treatment,simulation and ergonomic design for different individuals,the existing digital anatomical atlases of the head are difficult to accurately express the anatomical morphological characteristics of different individuals due to the lack of anatomical morphological differences between individuals,thus restricting their application in personalized modeling.Considering this problem,the digital anatomical atlases of head with personalized deformation are constructed based on the CT image dataset of Chinese population.In the aspect of atlas construction method,an atlas construction method without image segmentation is proposed,which incorporates the information of anatomical landmarks.In the aspect of atlas application,an automated brain parcellation algorithm of PET/CT images based on personalized atlas registration is proposed.The main contributions are listed as follows.(1)In view of the lack of anatomical differences among individuals in the existing digital head anatomical atlas,the deformable statistical atlases were constructed based on the CT images of 65 healthy Chinese adults.A thin-plate spline robust point matching with anatomical landmark constrained algorithm is proposed to realize the accurate registration of head reference template to CT image dataset,and obtain the shape correspondence between training samples.Furthermore,the statistical shape modeling algorithm is used to construct the deformable statistical atlases of male and female heads,which includes the inter-subject anatomical morphological variations of the population.The experimental results show that the distribution of cephalometric measurements of the atlas are consistent with the existing large-scale adult Chinese statistics,which the Zs indices of the somatometry are in the range of[-1.96,1.96],and the RR indices of the osteometry are in the range of[-0.2,0.2].The deformable statistical atlases of the head not only provide complete and high-precision 3D anatomical structure,but also can accurately deform to different subjects to realize personalized anatomical modeling and meet the needs of personalized design and diagnosis and treatment.(2)In the process of constructing deformable statistical atlas,image segmentation of the training dataset require a lot of manual operation,which is especially serious for the head with multiple anatomical structures.To solve this problem,a merged template registration algorithm based on anatomical landmarks and imaging intensity information,which uses manually marked key anatomical landmarks instead of manual image segmentation to reduce the manual interaction workload of statistical atlas construction method.The experimental results show that this method achieves pixel-level shape correspondence accuracy(1.38 mm)for skull and sub-pixellevel accuracy(0.82 mm)for skin.The method not only simplifies the manual interaction,but also ensures the template matching accuracy effectively.In addition,the presented method can also be extended to other parts of the head that contain many anatomical structures.In addition,the method can also be extended to other parts of the body especially contain multiple anatomical structures.An example of the construction of a deformable statistical atlas of the spine is shown,and the template matching accuracy can be achieved at the sub-pixel level(0.92 mm).(3)Based on the implementation of atlas construction,this paper preliminarily explored the application of atlas in personalized diagnosis and treatment.An automated brain parcellation algorithm based on deformable statistical atlas of head in PET/CT images is proposed.Because most of the existing atlas registration algorithms adopt global optimization strategy,they cannot effectively optimize the registration accuracy of local brain structures.An automated algorithm is proposed to detect the key anatomical landmarks in deep brain regions of PET images based on deep Q network to constrain local small-scale brain structure registration.Furthermore,a strategy combining the characteristics of PET and CT dual-modality images is proposed to improve the accuracy of head global-scale registration.The proposed method takes into account both global-scale and local small-scale registration,and realizes the automated and accurate segmentation of brain volume of interest.Compared with the existing atlas-based segmentation methods,the proposed method can achieve more accurate segmentation of brain regions.The average similarity coefficient is 79.39%,the average surface distance is 0.97 mm at the PET sub-pixel level,and the volume recovery ratio close to 1.The atlas constructed in this study has potential applications in personalized head simulation,plastic surgery,head radiotherapy target delineation,and brain disease diagnosis and treatment,etc.The atlas constructed in this study provides reference atlas data and personalized modeling methods for subsequent application studies.