JOINT LOCAL AND NON-LOCAL IMAGE REGULARIZATION FOR MODEL BASED ITERATIVE RECONSTRUCTION
Publication Date: 2016-May-23
The IP.com Prior Art Database
A technique utilizing both local and non-local regularization constraints for computed tomography (CT) image reconstruction is disclosed. The technique includes a hybrid reconstruction algorithm that combines model-based iterative reconstruction (MBIR) with conventional neighboring pixel based local image regularization, such as total variation (TV) image regularization and Dictionary Learning (DL) based non-local method (MBIR-DLTV). According to an embodiment of the invention, a new version of Ordered-Subsets Separable Palaboloidal Surrogate (OS-SPS) method is proposed for DL based reconstruction. The OS-SPS method provides an efficient implementation of cost optimization. In the new version of OS-SPS method, the dictionary is updated less frequently than the image itself.
The present invention relates generally to X-ray computed tomography (CT) and more particularly to a technique that utilizes advantages from both local and non-local image regularization methods for improving image quality obtained from X-ray CT.
Generally, luggage inspection at airports is performed using transportation security X-ray computed tomography (CT) scanners. The security luggage CT reconstruction has different challenges from medical CT reconstruction. For example, luggage used during transportation generally includes highly attenuating materials, such as, metals and glasses. Image reconstruction of such materials typically suffers severe streaking artifacts due to photon starvation and low signal-to-noise ratio. Consequently, the scanner is unable to detect hazardous materials. A higher energy ray is usually used for better penetration than medical CT scanners. However, total amount of dose is required to be kept low to protect a scanning tube.
Model based iterative reconstruction (MBIR) is a statistical reconstruction method which can control a tradeoff between data fitting and image regularization by optimization of a cost function that integrates system data acquisition physics, photon statistics, and prior knowledge of reconstructed volume. The regularization term has a role of controlling image attributes, such as, smoothness and edge preservation level. A conventional regularization term is based on pair-wise neighboring pixels in N-point neighborhood system, and typically a function of pixel value difference between the two neighboring pixels.
However, use of neighboring pixel pair interaction is sometimes too local and too simple. Also, such neighboring interaction is not effective to restore object structures and boundaries. For example, for transportation security computed tomography (CT) application, robust performance of image segmentation or detection algorithm in post processing is critical. As a result, implementation of image regularization in a non-local way with use of wider range of pixel interactions is preferred over neighboring pixel pair interactions.
Dictionary learning (DL) is a technique which was originally introduced as a successful denoising tool in image processing. In DL, signals are fitted to a sparse linear combination of dictionary atoms which are predefined. The basic idea of DL is to select a few elements from over complete dictionary which can be built either offline (global DL) or online (adaptive DL). As DL shows outstanding performance in denoising applications, DL is applied to CT reconstruction.
Both DL and conventional neighboring pixel based regularization have advantages and disadvantages. The use of neighboring pixel interaction is sometimes too local and too simple, and is not effective to restore object structures and boundaries. But it tends to generate consistent results. DL uses n...