[Paper Review] DECT-MULTRA: Dual-Energy CT Image Decomposition With Learned Mixed Material Models and Efficient Clustering

Zhipeng Li, Saiprasad Ravishankar+2 more|arXiv (Cornell University)|Jan 1, 2019

Advanced X-ray and CT Imaging26 references4 citations

TL;DR

This paper proposes DECT-MULTRA, a novel image-domain dual-energy CT decomposition method that combines penalized weighted-least squares (PWLS) with a learned mixed union of sparsifying transforms (MULTRA) to improve material image quality. By pre-learned common- and cross-material transforms, the method enables efficient, closed-form optimization that reduces noise and artifacts while preserving edges and fine structures, outperforming existing methods on both XCAT phantom and clinical head data.

ABSTRACT

Dual energy computed tomography (DECT) imaging plays an important role in advanced imaging applications due to its material decomposition capability. Image-domain decomposition operates directly on CT images using linear matrix inversion, but the decomposed material images can be severely degraded by noise and artifacts. This paper proposes a new method dubbed DECT-MULTRA for image-domain DECT material decomposition that combines conventional penalized weighted-least squares (PWLS) estimation with regularization based on a mixed union of learned transforms (MULTRA) model. Our proposed approach pre-learns a union of common-material sparsifying transforms from patches extracted from all the basis materials, and a union of cross-material sparsifying transforms from multi-material patches. The common-material transforms capture the common properties among different material images, while the cross-material transforms capture the cross-dependencies. The proposed PWLS formulation is optimized efficiently by alternating between an image update step and a sparse coding and clustering step, with both of these steps having closed-form solutions. The effectiveness of our method is validated with both XCAT phantom and clinical head data. The results demonstrate that our proposed method provides superior material image quality and decomposition accuracy compared to other competing methods.

Motivation & Objective

• To address the limitations of conventional image-domain DECT decomposition, which suffers from noise amplification and artifacts due to sensitivity to image noise.
• To improve material decomposition accuracy and image quality by incorporating learned prior knowledge from material patches.
• To develop a computationally efficient method that combines sparsity-based regularization with adaptive clustering of image patches to the best-matching sparsifying transforms.
• To generalize the MULTRA model trained on phantom data to real clinical DECT data, demonstrating robustness and transferability.
• To outperform existing state-of-the-art methods such as DECT-EP, DECT-TDL, and DECT-ST in both noise suppression and edge preservation.

Proposed method

• The method uses a penalized weighted-least squares (PWLS) formulation for image-domain DECT decomposition, incorporating regularization via a learned mixed union of sparsifying transforms (MULTRA).
• Common-material sparsifying transforms are pre-learned from patches of individual basis materials (e.g., water, bone), capturing shared structural properties across material images.
• Cross-material sparsifying transforms are learned from multi-material patches to model interdependencies between different materials.
• Image patches are adaptively assigned to the best-matching sparsifying transform via clustering during optimization, enabling flexible and data-driven sparsity representation.
• The optimization alternates between an image update step and a sparse coding and clustering step, both of which have closed-form solutions, ensuring computational efficiency.
• The method is initialized using Direct Matrix Inversion and further refined using DECT-EP and DECT-TDL results, improving convergence and stability.

Experimental results

Research questions

• RQ1 Can a learned mixed union of sparsifying transforms effectively model both common and cross-material dependencies in DECT image decomposition?
• RQ2 Does the proposed MULTRA-based regularization significantly reduce noise and artifacts compared to non-adaptive and dictionary-based methods in image-domain DECT?
• RQ3 Can a model pre-trained on a digital phantom generalize effectively to real clinical DECT data without retraining?
• RQ4 How does the performance of DECT-MULTRA compare to DECT-EP, DECT-TDL, and DECT-ST in preserving fine anatomical structures such as small vessels and marrow details?
• RQ5 Can the method achieve high-quality decomposition with low computational cost while maintaining robustness to image noise?

Key findings

• DECT-MULTRA significantly reduces noise and artifacts at material boundaries compared to Direct Matrix Inversion and DECT-EP, particularly in low-contrast regions.
• The method successfully preserves a small dark spot in the water image corresponding to a diluted iodine-containing artery, which was missed by DECT-EP and Direct Matrix Inversion.
• DECT-MULTRA produces sharper edges in soft tissue structures, such as marrow, compared to DECT-TDL, as confirmed in zoomed-in regions of the clinical data.
• The MULTRA model trained on the XCAT phantom generalized well to clinical head DECT data, demonstrating strong transferability and robustness.
• Quantitative evaluation showed that DECT-MULTRA outperformed DECT-EP, DECT-TDL, and DECT-ST in both noise suppression and structural preservation, with superior decomposition accuracy.
• The closed-form optimization steps enabled efficient computation, making the method suitable for practical clinical deployment despite its advanced regularization framework.