X-ray computed tomography (CT) is now a widely used clinical imaging modality for screening and diagnosis, emergency medicine, image-guided interventions, and monitoring of therapeutic responses with its high spatial resolution, short scanning time and low patient cost. But the excessive X-ray radiation may induce cancer, leukemia, or increase biological risk, so radiation doses associated with CT have received serious concern. Among all the methods proposed so far to obtain low radiation doses, the most directly and practical used method is to lower the X-ray tube current, however reduce tube currents make the projection data corrupted by quantum noise. Meanwhile, the associated reconstructed image usually suffers from serious noise and streak artifacts, which has negative influence for clinical diagnosis and analysis. Therefore, how to improve the quality of reconstruction image under the lower radiation dose CT scan has been the difficult problem in CT research field, and has the important scientific research value and the clinical use value.
As above, to overcome quality degradation of CT image caused by low X-ray tube current scanner protocols, this dissertation focuses on incorporating suitable feature sparse representation into the statistical iterative reconstruction, CT image processing and projection data restoration frameworks for better performance of LDCT imaging. The main innovative works are summarized as follows:
(1) 3D feature constrained reconstruction for LDCT imaging.
A great deal of progress has been made for the study of signal/image processing based on feature dictionary learning (DL) and sparse representation in recent years. In this dissertation, inspired by the idea of high-dimensional feature dictionary and sparse representation techniques, we introduced them into the statistical iterative reconstruction scheme and constructed a 3D feature constrained reconstruction (3D-FCR) algorithm for LDCT. First, the available high quality standard-dose CT (SDCT) images which can serve as sample data so as to constructed 3D tissue feature dictionary by feature extraction and training. Then, rich 3D anatomy tissue features in the forms of dictionary atoms were incorporated as constraints to overcome the ill-posedness of reconstruction with noisy measurements in CT, which is of great benefit to image update. At last, an alternating iterative estimation scheme was given. Experiments on both phantom and clinical data were conducted to validate the performance of this method. Compared with the classic reconstruction methods, the proposed method has the excellent performance in protection of the tissue detail and removal of streak artifact.
(2) Discriminative feature representation for LDCT image processing.
Successful application of DL method has been reported in LDCT image processing. It is often the case that strong noise and streak artifacts (with strong intensities) can also be sparsely representable and as a result cannot be removed effectively. To overcome this problem, we further explored the improvement of CT image feature representation ability and proposed a novel method called the discriminative feature representation (DFR). This DFR method assumes LDCT images as the superposition of desirable anatomical component and undesirable noise-artifact component induced by low dose scan protocols, and the decomposed anatomical component was used to provide the processed LDCT images with higher quality. The significant morphological distinction between anatomical structure and noise-artifact structure is leveraged in algorithm devising. The target anatomical component was solved via the DFR algorithm using a featured dictionary composed by atoms representing anatomical features and noise-artifact features. Comparative experiments with abdomen LDCT data validated the good performance of the proposed approach, and it will have good clinical application prospects.
(3) Discriminative prior - prior image constrained compressed sensing reconstruction for LDCT imaging.
Based on discriminative feature representation and prior image constrained compressed sensing reconstruction model, a joint estimation framework termed discriminative prior - prior image constrained compressed sensing (DP-PICCS) reconstruction was proposed. To address the mismatch problem between the prior image and target image for the DP-PICCS algorithm, this article applied DFR method to obtain high quality image volume by processing the FDK reconstruction volume. Additional, the DP-PICCS model utilizes discriminative prior knowledge via two feature dictionary discriminative constraints which built on atoms from the samples of anatomical structure feature patches and noise-artifacts residual feature patches, respectively. The new model overcomes the mismatch problem in current PICCS reconstruction, and adopt discriminative prior constraints to improve reconstruction image quality. Experimental results shown that the DP-PICCS method can lead to a promising improvement in terms of the suppressing noise-artifacts and preserving details structures.
(4) Discriminative feature representation to improve projection data inconsistency for LDCT imaging.
Directly using measured noisy projections to reconstruction can significantly deteriorate images due to the data inconsistency under LDCT scan protocols. The data consistency requirement imposed by fidelity term is intrinsically questionable due to the fact that the desirable reconstruction image reprojection always does not always match with the measured noisy projections in LDCT imaging. To deal with this problem, we proposed here a new sinogram restoration approach, the sinogram - discriminative feature representation (S-DFR) method, and the final improved LDCT images were reconstructed from restored projection data. First, the new approach works through a 3D representation based feature decomposition of the projected attenuation component and the noise component using a well-designed composite dictionary containing atoms with discriminative features. Then, the data consistency conditions (DCCs) were imposed on the decomposed projections to futher ensure data consistency. Its comparison to other competing methods through experiments on simulated, clinical and rat projection data demonstrated that the S-DFR method offers a sound alternative in LDCT imaging.