Improvement of thoracic hybrid PET/CT registration using hybrid feature with combined intensity multimodal demon with PET sinogram filtering

Accurately registered and fused PET/CT images are required for better tumor interpretation and the following tumor management in oncology and radiotherapy purposes. Although the hybrid PET/CT machine is supposedly solves the problem of misregistration between the PET/CT images, the offered solution is not optimal. The nonlinear misregistration due to physical and physiological motions stays on, declining the performance of the hybrid PET/CT machine. Therefore, the aim of this thesis is to solve the misregistration problem inflicting the PET/CT images acquired from the hybrid PET/CT scanner. Overall, the proposed registration method consists of three major steps. The first step is to perform 3D hybrid mean-median filtering based on the weighted average scheme on the PET sinogram domain. The second step is to segment selected structures of the thorax region which are the lungs, the heart and the tumor in both PET/CT images using a specific segmentation method for each structure excluding the heart in the PET image in which the segmentation is manually done. The main focus at this part is to design segmentation methods for the PET lung and the CT heart as these two subjects are rarely addressed. These segmented structures are used as “features” in the third stage where hybrid feature combined intensity multimodal demon registration is carried out to register both images. This method which is an improved version of multimodal demon registration uses a combination of mutual information (MI), sum of conditional variations (SCV) and multimodality independent neighborhood descriptive (MIND) similarity measures. The PET sinogram filter is tested on the NCAT based PET sinograms generated using ASIM PET simulator of different signal to noise ratio (SNR) and is compared with standard filter as used in analytical filtered-backprojection (FBP) reconstruction method. Aside from FBP, the improvement made by the filter on the iterative maximum likelihood expectation maximization with median root prior (MRP-MLEM) reconstruction method is also investigated. The filter significantly improves the global and local SNR of the PET image by more than 40% and more than 150% when compared to Hanning filtered FBP and MRP-MLEM reconstructed images without filtering. In terms of contrast to noise ratio (CNR), the proposed filter constantly generates improved CNR for all datasets in both analytical and statistical reconstruction methods. In the second stage, the proposed segmentation methods are evaluated on simulated NCAT based PET/CT and 21 clinical patient datasets. Apart from satisfactory subjective evaluation through visual displays,the segmentation of two structures, CT heart and PET lung are validated against expert segmentation on 10 datasets. The achieved mean Dice and Jaccard coefficients for both structures are more than 0.8. Then, the proposed improved intensity multimodal demon registration is tested on simple images and various types of medical images and the registration results are satisfactory. Specific to PET/CT registration problem, the proposed hybrid feature intensity multimodal registration method is tested on the simulated NCAT PET/CT images acquired at different breathing phases as well 21 clinical hybrid PET/CT datasets. Experimental results show that the combination of SCV and MIND based similarity measures produces the best registration result for PET/CT misregistration problem. In particular to clinical datasets experiment, the mean NMI improvement achieved by the proposed hybrid feature combined intensity multimodal demon registration is twice than the established free form deformation (FFD) registration method. The success of the registration of the patient datasets is also validated through improved lung volume overlap between the PET lung and the CT lung post registration according to Jaccard and Dice coefficients calculations. The registration method increases the Jaccard and Dice measures by 7.78% and 4.46% in average respectively after registration.

Preview Text FK 2016 59 - IR.pdf Download (1MB) | Preview

Actions (login required)

View Item