Image Reconstruction from Incomplete Projection of Medical Data
Supervisor(s) and Committee member(s): Dr Linganagouda Kulkarni (supervisor), BVB College of Engineering & Technology, Hubballi-580031, Karnataka State, India
Medical imaging has advanced tremendously over the decade right from the inception. Among many, X-ray Computed Tomography+ (CT) is recognized as an imperative medical imaging modality to reveal the interior details of human body for effective diagnosis, treatment, operation and complication management of various clinical cases. CT operates on the principle of reconstruction of an image from projection based on ‘Radon transform’. Here, reconstruction of image is a function of view angle, rotational increment and projections which are attenuation of X-ray. In many situations, data acquired by CT could be incomplete in nature, which lead to the addition of noise, artifact and eventually degradation of image quality. The aim of this research work is to improve diagnostic quality and acceptability of reconstructed images, by addressing effects of incomplete data problem, which arises in three situations a) truncated projections b) limited view and c) sparse view.
Truncated projections: Metal-induced artifacts are one of the most common streaking artifacts# in CT images which are caused by metallic implants within the patients’ body. These streaking artifacts are due to attenuation of X-ray photons passing through these objects, resulting in truncated projections, where gaps are formed in projection data. Metal artifacts can corrupt CT images such that they become challenging to read and interpret, eventually leading to limited diagnostic value. In this regard, two new metal artifact reduction (MAR) methods for addressing truncated projection problem are proposed. First MAR is a novel 3-Step method, where CT slices which are adjacent to artifact affected are used to correct metal artifact affected CT image. It is a hybrid approach involving the sinogram interpolation correction, projection-profile prediction and iterative algebraic reconstruction technique (ART). Second MAR method is a ‘user intervention based metal segmentation’ using ilastik© software. Segmentation of metal part is carried out by two methods i) Random forest classifier (RFC) based train and testing method and ii) Seeded watershed segmentation (SWS) method. The proposed MAR methods are compared with recently reported iterative frequency split–normalized (IFS) and L0 – Douglas-Rachford splitting (L0-DRS) MAR approaches using clinical cases of hip prostheses, maintaining the similar experimental environment. Performance comparison by means of ‘reference free ground truth’ metric revealed greater ability and acceptability of proposed methods. Among two proposed MAR methods; 3-Step MAR method emerged as most promising technique, with excellent quantitative and qualitative score.
Limited view: Digital breast tomosynthesis (DBT) is an emerging imaging modality which produces 3-D radiographic images of the breast. However DBT reconstructs tomographic images from a limited view angle, thus data acquired from DBT is not sufficient enough to reconstruct an exact image. The image constructed at abstract plane for user defined depth and inclination is observed to have poor in-plane resolution yielding artifacts. In this thesis, tomosynthesis system performance is optimized with respect to angular range and number of projections. A New Shift and Magnified Projections Algorithm (SMP) is developed to construct image at abstract plane, and evaluated by using in-house developed 3-D Breast phantom model. XRaySim© simulator is employed for simulating DBT acquisition system. Performance of proposed method is compared with recently published i) α-trimmed, principle component analysis (PCA)
based back projection (BP) method and ii) algebraic reconstruction with 3-D total variation ( ART +TV3D) based reconstruction method. During comparative study, common computational platform and environment are maintained. Performance evaluation by quality metrics such as artifact spread function (ASF), contrast to noise ratio (CNR) and structural similarity index (SSIM) demonstrated better results, in addition to speed-up in computation time.
Sparse view: CT benefits clinical decisions, when used for appropriate indications, however, concerns have been raised regarding the potential risk of cancer induction from repeated usage of CT. Keeping radiation dose as low as reasonably achievable (ALARA), consistent with the diagnostic task, is the main strategy for decreasing this potential risk. One of the strategy to reduce the radiation dose, is to acquire the projections at sparse rotational angle. This eventually reduces number of projections and thereby radiation dose. Due to undersampling, streaking artifacts will be observed in the final image, which becomes prominent as sampling is reduced. To address this issue an optimized ‘Selective highly constrained back-projection’ (S-HYPR) method is developed, and experimented using cardiac CT Angiogram (CTA) clinical cases. Performance of S-HYPR is compared with recently reported iterative reconstruction (IR) and adaptive iterative dose reduction in 3-D (AIDR3D) method. S-HYPR emerges as most promising low dose imaging method, when assessed through quality metrics such as ASF, CNR and SSIM.
A huge computation time demand by CT image reconstruction Filter Back Projection (FBP) algorithm is addressed by Compute Unified Device Architecture (CUDA) based graphics processor unit (GPU), a parallel processing model. The forward and backward projection algorithms, which are basic building blocks of the image reconstruction process are programmed to execute in pipeline structure, by exploring the libraries of CUDA. Considerable acceleration factor is recorded, compared to a sequential, single thread and multi-threaded CPU implementation. During this experiment, quality of reconstructed image is observed to be consistent, with improved speed-up in computation time. The usage of GPU in medical imaging applications alleviates huge computational demand of complex CT image reconstruction processes.
Incomplete data problems observed in three discrete situations as cited above are of great relevance and are current research issues in the health-care community.