Image domain dual material decomposition for dual‐energy CT using butterfly network

Abstract

Purpose Dual‐energy CT (DECT) has been increasingly used in imaging applications because of its capability for material differentiation. However, material decomposition suffers from magnified noise from two CT images of independent scans, leading to severe degradation of image quality. Existing algorithms exhibit suboptimal decomposition performance because they fail to fully depict the mapping relationship between DECT images and basis materials under noisy conditions. Convolutional neural network exhibits great promise in the modeling of data coupling and has recently become an important technique in medical imaging application. Inspired by its impressive potential, we developed a new Butterfly network to perform the image domain dual material decomposition. Methods The Butterfly network is derived from the model of image domain DECT decomposition by exploring the geometric relationship between the mapping functions of the data model and network components. The network is designed as the double‐entry double‐out crossover architecture based on the decomposition formulation. It enters a pair of dual‐energy images as inputs and defines the ground true decomposed images as each label. The crossover architecture, which plays an important role in material decomposition, is designed to implement the information exchange between the two material generation pathways in the network. The proposed network is further applied on the digital phantom and clinical data to evaluate its performance. Results The qualitative and quantitative evaluations of the material decomposition of digital phantoms and clinical data indicate that the proposed network outperforms its counterparts. For the digital phantom, the proposed network reduces the standard deviation (SD) of noise in tissue, bone, and mixture regions by an average of 95.75% and 86.58% compared with the direct matrix inversion and the conventional iterative method, respectively. The line profiles and image biases of the decomposition results of digital phantom indicate that the proposed network provides the decomposition results closest to the ground truth. The proposed network reduces the SD of the noise in decomposed images of clinical head data by over 90% and 80% compared with the direct matrix inversion and conventional iterative method, respectively. As the modulation transfer function decreases to 50%, the proposed network increases the spatial resolution by average factors of 1.34 and 1.17 compared with the direct matrix inversion and conventional iterative methods, respectively. The proposed network is further applied to the clinical abdomen data. Among the three methods, the proposed method received the highest score from six radiologists in the visual inspection of noise suppression in the clinical data. Conclusions We develop a model‐based Butterfly network to perform image domain material decomposition for DECT. The decomposition results of digital phantom validate its capability of decomposing two basis materials from DECT images. The proposed approach also le...

Physics-informed Deep Learning for Dual-Energy Computed Tomography Image Processing

“…Our results are consistent with, and supplement a prior study showing that convolutional networks can provide excellent performance in noise suppression on DECT images. Zhang et al 18 . reported a remarkable, up to 95%, reduction in noise standard deviation in tissue, bone, and mixture regions on a digital phantom reconstructed from dual-energy projections.…”

Deep learning based spectral extrapolation for dual‐source, dual‐energy x‐ray computed tomography

“…Second, our current approach to spectral extrapolation visibly smooths the chain B image data outside of the training mask (Figs. [5][6][10][11][12]. This smoothing is a result of matching independent noise observations between the chain A data and the known chain B data used as training labels.…”

“…For rods within the training mask (1-9), the CNN explicitly preserves the chain B contrast (maximum error by magnitude, 0.33 HU). For rods outside of the training mask and in the training halo (10)(11)(12)(13)(14)(15), however, the CNN spectral extrapolation results are less consistent, particularly for the 2 and 15 mg/mL iodine rods (contrast estimation errors: −17.34 and −28.89 HU, respectively). Compared with the much smaller average errors seen in the testing data (Fig.…”

“…8 This includes a growing body of work on CNN-based postreconstruction denoising for reproducing full-dose CT data from low-dose CT data. [9][10][11] More relevant here, this also includes several papers demonstrating the advantages of spatial contextual information in the problems of material decomposition [12][13][14] and the synthesis of noncontrast images 15 and monochromatic images. 16 More broadly, application of deep learning to the problem of spectral extrapolation is motivated by prior successes in a number of closely related image processing problems, including inpainting, 17 pan-sharpening, 18 and assigning color to grayscale images.…”

Fully connected neural network for virtual monochromatic imaging in spectral computed tomography

“…Liao et al 16 explored the feasibility of neural network in obtaining material decomposition image using single-energy CT and verified it experimentally with clinical patient data. Zhang et al 17 proposed the Butterfly network to implement material decomposition in an image domain. They verified that the Butterfly network yielded excellent performance in image quality improvement and noise suppression.…”