Deep-learning-based direct synthesis of low-energy virtual monoenergetic images with multi-energy CT

A multienergy computed tomography method without image segmentation or prior knowledge of X-ray spectra or materials

“…This method is based on a blind source separation variant of multivoltage projections with a stepped voltage potential scan. Compared with multienergy CT methods that require photon-counting detectors, such as the methods of references [25] , [26] , [27] , [29] , [30] , [33] , this method requires no hardware changes in traditional CT systems except voltage control. Compared with most multienergy CT methods based on traditional CT systems, the proposed method has four advantages.…”

“…Xiaochuan Wu et al improved fully convolutional DenseNets for multimaterial decomposition [29] . Hao Gong et al developed a deep-learning method for the direct synthesis of low-energy virtual monoenergetic images [30] . Weiwen Wu et al developed a U-net approach for image reconstruction with -norm, total variation, residual learning, and anisotropic adaptation [31] .…”

Deep learning‐based virtual noncalcium imaging in multiple myeloma using dual‐energy CT

“…$${B_{CNN,k,n}}$$ denotes the pixel‐wise binary material‐specific mask generated from the classification branch. $${L_{IGC}}$$ denotes the image gradient correlation based regularizer 28,31 that ensured edge consistency in both material domain and DECT image domain, $$\nabla ( \cdot )$$ denoted the anisotropic form of image gradient, $$\rho ( \cdot )$$ denoted Pearson correlation, and ε was a small constant (fixed at $$1.0\, \times \,{10^{ - 4}}$$). $${L_{Feat}}$$ is the feature reconstruction loss 33 that ensured high‐level texture consistency between the mixed images $${I_{CNN,mix}}$$ and $${I_{DECT,\;mix}}$$ by matching the high‐level features extracted from the l th convolutional layer (empirically fixed at the 15th layer) of a pretrained VGG‐19 neural network $$\phi ( \cdot )$$, $${I}_{\textit{CNN},k,\textit{mix}}$$ and $${I}_{\textit{DECT},k,\textit{mix}}$$ were the mixed images from $${I_{CNN}}$$ and $${I_{DECT}}$$…”

“…Hyperparameters were all empirically determined from prior studies. Briefly, the setup of stem CNN was determined using our prior setup in references, 28,31 and the setup of bifurcated branches was determined in our recent study. 35 The hyper-parameters (e.g., weighting factors) in loss function were also similarly set up.…”

“…B CNN,k,n denotes the pixel-wise binary material-specific mask generated from the classification branch. L IGC denotes the image gradient correlation based regularizer 28,31 that ensured edge consistency in both material domain and DECT image domain, ∇(⋅) denoted the anisotropic form of image gradient, 𝜌(⋅) denoted Pearson correlation, and 𝜀 was a small constant (fixed at 1.0 × 10 −4 ). L Feat is the feature reconstruction loss 33 that ensured high-level texture consistency between the mixed images I CNN,mix and I DECT, mix by matching the high-level features extracted from the l th convolutional layer (empirically fixed at the 15th layer) of a pretrained VGG-19 neural network 𝜙(⋅), I CNN,k,mix and I DECT,k,mix were the mixed images from I CNN and I DECT respectively, and 𝜆 was empirically fixed at 1.0 × 10 −3 .…”