Model-Based Image Processing Algorithms for CT Image Reconstruction, Artifact Reduction and Segmentation
Model-based image processing is a collection of techniques that provides a systematic framework for solving inverse problems in imaging systems. In this dissertation, three problems that arise in CT imaging systems are addressed using the model- based approach: image reconstruction for the single energy X-ray CT with both 2D parallel-beam and 3D multi-slice geometries, simultaneous image reconstruction and beam hardening correction for the single energy X-ray CT, and simultaneous metal artifact reduction and image segmentation for CT images. In the first topic, the methodology of model-based iterative reconstruction (MBIR) for solving CT image reconstruction problems is studied. Recent research indicates that the MRIR has potential to improve image quality and remove artifacts comparing to traditional filtered back-projection (FBP) methods. The MBIR algorithms for both 2D parallel-beam and 3D multi-slice helical CT geometries are developed using the formulation under the statistical framework and the reconstruction is solved using optimization techniques. The result on the real CT baggage dataset is presented, which illustrates the image quality improvement and noise and artifact reduction. The second topic studies the beam hardening correction problem in the single-energy X-ray CT. Beam hardening is the effect that material preferably attenuates more low-energy X-ray than high-energy, and with a broad X-ray source spectrum, assumption that distinct materials can be separated according to their densities, a more accurate forward model that accounts for the X-ray spectrum is developed and a MBIR algorithm that incorporates this new model is proposed. The overall algorithms works by alternating estimation of the image and the unknown model parameters, therefore no additional information is required. Results on both the simulated and real CT scan data show that the proposed method significantly reduces metal streak artifacts in the reconstruction. The third problem is the segmentation of CT images with metal artifacts and without the access to the CT data. Segmenting interesting objects from CT images has a wide range of applications in medical diagnosis and security inspection. How- ever, raw CT images often contain artifacts such as streak due to the dense metal objects, and these artifacts can make accurate segmentation difficult. A novel model- based approach that jointly estimates both the segmentation and the restored image is proposed and the unified cost function consists of three terms: 1) a data fidelity term that relates the raw and restored image and incorporates a streak mask; 2) a dictionary-based image prior which regularizes the restored image; 3) a term based on the continuous-relaxed Potts model which couples the restored image intensities and segmentation labels. Results on both simulated and real CT data are presented and support that the joint segmentation and MAR can produce superior results without the use of the raw CT data.
Bouman, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our