Non-rigid registration for multi-modality image fusion using prior shapes

Chronic obstructive pulmonary disease (COPD) is a chronic lung disease that causes breathing difficulties. One possible course of treatment for severe COPD is lung volume reduction surgery (LVRS), which involves removing, or isolating, the lobe or lobes of the lung that are most affected by the disease. A fusion of the multi-slice computed tomography (MSCT) and ventilation (V) and perfusion (Q) single photon emission computed tomography (SPECT) modalities therefore represents a powerful tool to for COPD analysis and then for guiding the lung resection surgery.Due to reduced uptake of radioisotope at the location of lesion, the V and Q of a moderate COPD patient delineate photopenic regions, which are normally misrecognised as part of the background in the target SPECT scan. Non-rigid registration, which lacks displacement constraints, is therefore performed on MSCT scans with excessive deformations. Moreover, considering the low-resolution nature of functional imaging and highly deformable property of lungs, very few published algorithms are able to accommodate current clinical demands. The motivation of this project is to develop a high-performance, statistical deformation model (SDM)-based non-rigid registration algorithm capable of achieving accurate alignment of lung MSCT and SPECT imaging.In this project, an innovative similarity registration method for volumetric shapes is proposed at the beginning. The method is based on the characteristic function, and intended to strike a desirable balance between performance and efficiency. Radial moments and spherical coordinate system-based cross-correlation are exploited here to obtain the optimal scaling, rotation and translation parameters within a reasonable time. Moreover, an iterative method is also employed to improve the robustness of the algorithm. Group shapes in the presence of significant noise and lung shapes extracted from a low-dose computed tomography database are employed in the validation experiments.In order to eliminate the influence of the weighting parameter for the statistical term, a novel MSCT/SPECT registration technique based on a parameter-reduced SDM is proposed in this thesis. The SDM is trained on prior lung shapes. In addition, the multichannel technique performs V/MSCT and Q/MSCT alignments simultaneously to derive the optimal deformations. Lung MSCT and SPECT imaging data from a real medical database, as well as the 4D extended cardiac-torso phantom, were employed in the experiments. The algorithm proposed here was validated to be capable of preventing excessive deformations, and of achieving accurate registration between the two imaging modalities. The deformations for MSCT/SPECT registration are finally used to warp lobe masks, which are then mapped onto SPECT images for lung lobe/SPECT fusion