Learning-based multi-material CBCT image reconstruction with ultra-slow kV switching

Abstract

Objective

The purpose of this study is to perform multiple ( $\geq 3$ ) material decomposition with deep learning method for spectral cone-beam CT (CBCT) imaging based on ultra-slow kV switching.

Approach

In this work, a novel deep neural network called SkV-Net is developed to reconstruct multiple material density images from the ultra-sparse spectral CBCT projections acquired using the ultra-slow kV switching technique. In particular, the SkV-Net has a backbone structure of U-Net, and a multi-head axial attention module is adopted to enlarge the perceptual field. It takes the CT images reconstructed from each kV as input, and output the basis material images automatically based on their energy-dependent attenuation characteristics. Numerical simulations and experimental studies are carried out to evaluate the performance of this new approach.

Main Results

It is demonstrated that the SkV-Net is able to generate four different material density images, i.e., fat, muscle, bone and iodine, from five spans of kV switched spectral projections. Physical experiments show that the decomposition errors of iodine and CaCl $_{2}$ are less than 6 $%$ , indicating high precision of this novel approach in distinguishing materials.

Significance

SkV-Net provides a promising multi-material decomposition approach for spectral CBCT imaging systems implemented with the ultra-slow kV switching scheme.

Keywords

slow kV switching spectral imaging multi-material decomposition deep neural network

Introduction

Cone-beam computed tomography (CBCT) is widely used in image-guided interventions and radiation therapy. Despite its significant advantages, such as providing high isotropic resolution volumetric images at relatively low radiation doses and cost, the imaging performance of current CBCT systems still lags behind that of diagnostic multi-detector CT (MDCT), particularly in terms of image quality and the ability to provide quantitative imaging information.¹ Spectral CT imaging offers attenuation information across different X-ray spectra, enabling material decomposition and improved tissue differentiation. In clinical practice, spectral CT has demonstrated promising capabilities in image contrast enhancement,² artifact mitigation,³ and material recognition.^4,5

There are three major potential approaches to achieve spectral CBCT imaging. The first is dual-source technique,^6,7 where two X-ray tubes with different energy settings are used simultaneously to acquire data at two different energies. This method allows for spectral imaging and material decomposition but suffers from the disadvantage of increased system complexity and higher cost due to the need for two separate X-ray tubes, as well as potential cross scattering challenges. Another approach is the use of dual-layer detectors,^8,9 which consist of two layers of detector materials that capture X-ray data at different energies. This technique provides simultaneous dual-energy information, but it can lead to increased detector thickness and lower spatial resolution due to the dual-layer design. A third method is fast kV switching,^10,11 where the X-ray tube alternates between high and low kV settings during the scan, rapidly switching between different energy levels to capture spectral information. It has the disadvantage of longer scan times and also requires high-cost hardware.

Additionally, researchers are also working on developing new, cost-effective solutions for spectral CT technology. For instance, Petrongolo et al.¹² proposed obtaining high-energy and low-energy data by inserting a primary beam modulator between the X-ray source and the object. Tivnan et al.^13,14 introduced the use of a spatial spectral filter to generate multiple spectral channels on the detector surface. Another approach is the slow kV modulation technique, in which the X-ray tube voltage is gradually switched as the gantry rotates around the object.^1516–17 Compared to fast kV switching technology, the spectral information acquired through slow kV modulation is angularly sparse, which introduces additional challenges in material-specific CT image reconstruction. Therefore, specialized algorithms are required to achieve high-quality slow kV switching spectral CT imaging. For example, Szczykutowicz et al. proposed an iterative two-material decomposition algorithm based on the prior image constrained compressed sensing (PICCS) algorithm.¹⁸ Additionally, one-step material decomposition methods can also generate material-specific CT images by incorporating the slow kV modulation information into the forward imaging model. Notable examples include the joint statistical one-step iterative material image reconstruction algorithm proposed by Mechlem et al.,¹⁸ and the non-convex primal-dual one-step reconstruction algorithm proposed by Chen et al.¹⁹

Recently, deep learning techniques have been employed to enhance material decomposition in spectral CT imaging.^2021–22 Convolutional neural networks (CNNs) have been successfully applied to reconstruct two-material basis images in slow kV switching CT imaging.^23,24 While these approaches have been validated, it is important to note that CNNs have primarily been used for two-material decomposition. To the best of our knowledge, deep learning-based multi-material ( $\geq$ 3) decomposition with slow kV switching spectral information remains an unexplored topic.

In this study, we introduce an innovative CNN-based network, named SkV-Net, developed to achieve high-quality four-material decomposition in ultra-slow kV switching spectral CBCT imaging, where the tube voltage changes slowly from low to high voltage during a single rotation. The SkV-Net takes CT images reconstructed from each kV segment as input and automatically generates four material-specific CT images with high precision. Numerical experimental results demonstrate that the proposed SkV-Net can accurately predict multi-material images with excellent accuracy and image quality.

The rest of this paper is organized as follows: Section II introduces the design of the SkV-Net network, Section III provides details of data preparation and network training, Sections IV and V present the experimental setup and results, respectively, and Section VI provides the discussions and a brief conclusion.

Method

Imaging model

The spectral CT based on fast kV switching and ultra-slow kV switching techniques are illustrated in Figure 1. In fast kV switching technique,²⁵ the low and high tube voltages (80 kV and 140 kV) are switched on every alternate projection view. In contrast, each rotation of the CT scan based on ultra-slow kV switch contains several equally distributed spans of different kV settings. In this study, it is assumed that five kV spans are equally distributed between 85 kV and 125 kV, see the bottom illustration in Figure 1.

Figure 1.

The fast kV switching scheme (top), and the ultra-slow kV switching technique (bottom). In the fast kV switching scheme, it is assumed that the kV switches between the minimum and maximum tube voltages. In the ultra-slow kV switching scheme, it is assumed that the kV switches for five times between the minimum and maximum tube voltages.

To perform material decomposition, the imaging model of ultra-slow kV switching CT need to be built. Assuming the attenuation coefficient $μ (E, \vec{x})$ of a given material is equal to the product of an energy-dependent mass attenuation coefficient $μ_{m b} (E)$ and a position-dependent density distribution $ρ_{b} (\vec{x})$ , several pre-selected basis materials ( $b$ ) can be linearly combined as follows:

μ (E, \vec{x}) = \sum_{b} μ_{m b} (E) ρ_{b} (\vec{x}) .

(1)

Let $N_{i, k}^{0}$ represent the incident X-ray intensity of the ray $i$ and spectrum $k$ , the intensity after passing through the scanned object can be represented as:

N_{i, k} = N_{i, k}^{0} \int_{0}^{E_{max}} s_{i k} (E) \exp (- \int_{t \in l_{i}} μ (E, \vec{x} (t)) d t),

(2)

where

s_{i k} (E)

denotes the normalized X-ray intensity,

\int_{t \in l_{i}} (\cdot)

and

\int_{0}^{E_{max}} (\cdot)

denotes the integrals along the path

l_{i}

of the

i^{th}

ray and the energy

E

. In a discrete form, assuming the density map of the object is represented by

ρ_{b, j}

with

j

referring to the pixel index, the approximate measurement value

{\hat{N}}_{i, k}

can be expressed as:

{\hat{N}}_{i, k} = N_{i, k}^{0} \sum_{e} s_{i, k, e} \exp (- \sum_{b, j} μ_{m b, e} a_{i, k, j} ρ_{b, j}),

(3)

where

a_{i, k, j}

denotes the intersection length of ray

i

and pixel

j

under the energy spectrum

k

. All elements

a_{i, k, j}

with different

i

and

j

constitute the system matrix

A_{k}

. For measurements of different energy spectra,

A_{k}

can be the same or different, depending on the data collection strategy. To convert sinogram data into CT images, the most widely used analytical filtered back-projection (FBP) reconstruction algorithm is used to pre-reconstruct the sinogram:

x_{k} = B P (g * S_{k})

(4)

Here,

x_{k}

represents the CT images collected under energy spectrum voltage

k

S_{k} = \log (N_{i, k} / N_{i, k}^{0})

denotes the sinogram of spectrum

k

g

denotes the sinogram filter (Ramp filter in this study).

B P (\cdot)

denotes the backprojection operation. The material decomposition procedure of the deep-learning based method in ultra-slow kV CT can be represented as follows:

{{\hat{ρ}}_{b}} = N T {x_{k}}

(5)

Here,

{\hat{ρ}}_{b}

is the predicted material-specific images,

N T

represents the proposed material decomposition network SkV-Net.

The SkV-Net

Recent research has demonstrated that deep neural networks based on the U-Net architecture exhibit exceptional performance in medical imaging tasks.²⁵ By integrating advanced attention mechanisms into the U-Net framework, image quality can be further enhanced.^26,27 Inspired by these advancements, we have developed an innovative deep neural network, SkV-Net, designed to perform material decomposition in ultra-slow kV switching CT. The architecture of SkV-Net is illustrated in Figure 2. The CBCT volume images are processed in a slice-by-slice manner by SkV-Net. It takes five CT image slices, reconstructed from the limited-view sinograms at each kV setting, as input and outputs four material-specific density images, including water/muscle, bone/CaCl $_{2}$ , iodine, and fat/oil. The backbone of SkV-Net is based on the U-Net structure, consisting of 23 convolutional layers, each using 3 $\times$ 3 filters with a stride of 1. To reduce computational burden without significantly compromising image spatial resolution, only two pooling layers are employed. Batch normalization is applied in each layer to enhance training stability and improve the network’s generalization capabilities. Skip connections are incorporated to integrate spectral features across different levels of the network. To further enhance performance, a multi-head axial attention module²⁸ is incorporated in the layer with the highest-level features. This module helps capture more global information and provides a broader receptive field. The detailed structure of the multi-head axial attention module is shown in Figure 3. Let the input to the attention module be denoted as $χ \in R^{h \times w \times d_{i n}}$ , where $h$ represents height, $w$ represents width, and $d_{i n}$ represents the number of input channels. The query, key, and value matrices ( $q$ , $k$ , and $v$ ) are calculated as follows:

q = χ W_{q}, k = χ W_{k}, v = χ W_{v}

(6)

where

W_{q}, W_{k} \in R^{h \times (d_{i n} \times d_{q})}

and

W_{v} \in R^{h \times (d_{i n} \times d_{o u t})}

are learnable attention weight matrices. An output sequence

Y \in R^{h \times w \times d_{o u t}}

is generated as

Y = A v = s o f t max (q k^{T} / \sqrt{d_{i n} / h e a d s}) v

(7)

The core of the axial attention mechanism is centred on an affinity relationship

x W_{q} {W_{k}}^{T} x^{T}

, which facilitates the global aggregation of values

v

. This method enables the network to capture context information that is highly relevant yet non-local across the entire feature map, going beyond the local relationships that convolutional operations typically capture. The overall architecture of SkV-Net is designed to be layer-stacked, enabling the gradual extraction and fusion of features from low to high levels.

Figure 2.

Architecture of the proposed SkV-Net. The backbone structure is based on the U-Net. At the highest-dimensional image level, an attention mechanism is applied for spectral signal feature extraction.

Figure 3.

Structure of the axial attention model. The input consists of deep spectral signal feature maps extracted by the U-Net network, and the output is the feature maps encoded by the multi-head axial attention mechanism. Embedding the axial attention model aids in the fusion of deep spectral signal features, thereby more effectively guiding the decomposition of base materials.

Data and network training

Data preparation

Due to the unavailability of an ultra-slow kV switching CT system, it is challenging to collect spectral CT data and their corresponding basis material images. Therefore, we generate simulated data for network training²² and evaluate the material decomposition performance of SkV-Net using both numerical and experimental data. The labels used during network training are the basis images generated through natural image processing, as illustrated in III A.1. The input data for training consists of CT images reconstructed from the ultra-slow kV switching sinogram, which is synthesized from the basis images using the spectral CT imaging physics, see III A.2.

Basis image

The basis image generation scheme is shown in Figure 4. First, a large amount of natural images are downloaded from the ImageNet database.²⁹ Then the pixel values of the R-channel of the RGB image are normalized to a range between 0 and 1, and the resulting image is denoted as $I_{R}$ . Afterwards, $I_{R}$ is downsampled to 64 $\times$ 64, and then upsampled to 512 $\times$ 512 to generate a blurred version that preserves only the key structural information of $I_{R}$ . Finally, the basis material mask $M_{b}$ is generated based upon thresholding of the blurred image. The procedure can be expressed as:

M_{b} = T_{b} (F (I_{R}))

(8)

where

F

denotes the blurring operation,

T_{b}

represents the thresholding operation for material

b

. Specifically, the pixels with values greater than 0.85 are identified as bone material, those between 0.3 and 0.85 are considered as water, and pixels with values ranging from 0.1 to 0.3 are allocated to the fat mask. In addition, considering that fat substance also exists in the bone marrow, 50

%

possibility of overlap was set in the central parts of the bone and fat bases. Additionally, the iodine mask is generated independently, which contains randomly positioned circles with different diameters from 12 mm to 27 mm. To obtain the final basis image

ρ_{b}

, the mask

M_{b}

for each material is multiplied pixelwisely with the corresponding density map whose pixel values are adjusted to a reasonable range:

ρ_{b} = P_{b} (I_{R}) \cdot M_{b}

(9)

where

P_{b}

denotes the pixel value adjustment operation. To reflect the realistic material density characteristics, the density ranges for soft-tissues, bone, fat and iodine are 1.0-1.2 g/cm

^{3}

, 0.6-2.2 g/cm

^{3}

, 0.925-0.970 g/cm

^{3}

, 6.0-20.0 mg/ml. respectively.

Figure 4.

Key steps of generating numerical basis images from a natural image. Specifically, $P_{b} (\cdot)$ , $F (\cdot)$ , and $T (\cdot)$ represent the pixel value adjustment, image blurring, and thresholding operations, respectively. $I_{R}$ denotes the R channel of the RGB image, $M_{b}$ denotes the material-specific mask, and $ρ_{b}$ denotes the final material density map.

Ultra-slow kV switching sinograms

The ultra-slow kV switching projections were generated according to equation 3. The normalized spectra $s_{i, k, e}$ under five tube voltage of 85 kV, 95 kV, 105 kV, 115 kV, and 125 kV (same to the experimental settings) were generated with the SpekCalc software.³⁰ All spectra used the same filtration: 1.5 mm aluminium (Al) and 0.4 mm copper (Cu). The beam intensity $N_{i, k}^{0}$ was set to $10^{5}$ . The mass attenuation coefficients for basis materials were collected from the NIST database.³¹ The distance from the source to the detector (SDD) was 1560.6 mm and the distance from the source to the rotation centre (SOD) was 1156.3 mm. Projection data were collected from a full 360°rotation with 0.4°angular interval. The detector has 1 $\times$ 700 pixels with a pixel dimension of 0.417 $\times$ 0.417 mm $^{2}$ . Poisson noise was added to the projection data, and a $1 \times 5$ Gaussian kernel filter with sigma 0.52 was applied along the detector row direction to simulate the blurring effect of the detector.

Network loss

During network training, the loss function for SkV-Net is the mean squared error (MSE) between the material density image predicted by the network, denoted as ${\hat{ρ}}_{b}$ , and the corresponding ground truth, denoted as $ρ_{b}$ :

L = \frac{1}{W \times H \times B} \sum_{b} ‖ {\hat{ρ}}_{b} - ρ_{b} ‖_{2}

(10)

where

W \times H

represents the total number of pixels in the images,

B

denotes the number of basis materials,

‖ \cdot ‖_{2}

denotes the

l_{2}

norm.

Training strategy

The training of the SkV-Net was performed in the Pytorch platform environment on a high-performance workstation equipped with an Intel Xeon^® Silver 4210R CPU and an Nvidia RTX A6000 GPU. The initial learning rate was set to $10^{- 5}$ , and the learning rate decreases by a factor of 0.98 every 1000 training steps. The network was trained by 200 epochs with 10,000 pairs of training data.

Physical experiments

The experimental ultra-slow kV switching data were acquired from a benchtop CT system, which was equipped with a medical-grade X-ray tube (G-242, VAREX, UT, USA) and a flat-panel detector (4343CB, VAREX, UT, USA),as shown in Figure 5 (c). The geometrical parameters are listed in Table 1. Five full CT scans were performed with different tube voltages (from 85kV to 125kV in 10kV increments) and beam filtration of 1.5 mm Al and 0.4 mm Cu. The ultra-slow kV switching data were synthesized from these five datasets by selecting the corresponding projection data at different views. Two samples were scanned in this study, a tube phantom and a pork specimen, as shown in Figure 5 (a) and (b). In the tube phantom, seven tubes (10 mm in diameter) of different substances were immersed in water (90 mm in diameter), including one tube of vegetable oil, three tubes of calcium chloride (CaCl $_{2}$ ) solution at concentrations of 100, 200, and 300 mg/ml, three tubes of iodine solution at concentrations of 5, 10, and 20mg/ml.

Figure 5.

(a) and (b) are the tube phantom and pork specimen used in the physical experiments, (c) shows the experimental CT imaging benchtop.

Table 1.

Key parameters used for experimental data acquisition.

Parameter	XCAT	Tube Phantom	Pork Specimen
Tube voltage(kV)	85, 95, 105, 115, 125
Tube current (mA)	12.5, 8, 6.3, 4.5, 3.6
Filtration (mm)	AL: 1.5, Cu: 0.4
SOD(mm)	1156.3
SDD(mm)	1560.6
Detector array	1024 $\times$ 1024
Detector pitch(mm $^{2}$ )	0.417 $\times$ 0.417
Image size	512 $\times$ 512
Pixel size (mm $^{2}$ )	0.35 $\times$ 0.35
Training dataset	ImageNet	Simulated Tube	ImageNet
Training data (pair)	10000	10000	10000
Testing dataset	XCAT	Tube phantom	Pork specimen
Optimizer	Adam	Adam	Adam

Results

Numerical simulation results

The decomposition results of the numerical XCAT phantom are shown in Figure 6. Images in different columns correspond to four material bases: bone, iodine, fat and muscle. We can observe that the proposed SkV-Net could generate basis images that are very close to the ground truth with relatively small residuals. Different materials are well separated while fine anatomical structures are preserved, see the zoomed-in region of interest (ROI) in the fat image. The quantification results of the iodine inserts are listed in Table 2. The decomposed iodine densities are highly consistent with the true values with less than 6 $%$ relative errors. In addition, the mean absolute error (MAE) and the structural similarity index (SSIM) are calculated and listed in Table 3. The MAE values are small and SSIM values are close to 1, indicating that the SkV-Net could generate accurate material densities.

Figure 6.

Decomposition results of the XCAT phantom. From top to bottom, they correspond to the ground truth, SkV-Net results, and residual images. The display windows for bone, iodine, fat, and muscle basis images are [0, 1.4] g/cm $^{3}$ , [0, 18] mg/ml, [0, 1] g/cm $^{3}$ , and [0, 1.2] g/cm $^{3}$ , respectively. The display windows for the corresponding residual images are [0, 0.1] g/cm $^{3}$ , [0, 2] mg/ml, [0, 0.1] g/cm $^{3}$ , and [0, 0.1] g/cm $^{3}$ , respectively. The scale-bar denotes 35 mm.

Table 2.

The measured mean values and standard deviations(unit: mg/ml) of the iodine concentrations for numerical XCAT phantom.

Truth(mg/ml)	8	10	12	14	16	18
SkV-Net	8.46 $\pm$ 0.23	9.97 $\pm$ 0.14	11.95 $\pm$ 0.14	13.77 $\pm$ 0.15	15.50 $\pm$ 0.21	17.74 $\pm$ 0.18
Error( $%$ )	5.75	$- 0.8$	$- 0.42$	$- 1.64$	$- 3.12$	1.44

Table 3.

The measured MAE and SSIM values for numerical XCAT phantom.

Basis	MAE(mg/cm $^{3}$ )				SSIM
Basis	Bone	Iodine	Fat	Muscle	Bone	Iodine	Fat	Muscle
SkV-Net	1.0	0.02	2.4	3.4	0.9950	0.9892	0.9874	0.9887

Figure 7 demonstrate the CT images of different tube voltages synthesized from the decomposed material-specific images. The ground truth was generated via FBP reconstruction from the full scan projection data. We can observe that the CT images synthesized by the SkV-Net is highly consistent with the ground truth in terms of structural similarity and quantitative accuracy. These results indicates that high-quality CT images of different tube voltages can be obtained by the SkV-Net even though each voltage is only used in a limited-view angular range.

Figure 7.

Images from (a1) to (a5) are CT images reconstructed from the full scan projections of each kV. Images from (b2) to (b5) are CT images of different tube voltages synthesized from the decomposition results of SkV. All display Windows are [0, 0.42] cm⁻¹. The scale-bar denotes 35 mm. Plots in (c) and (d) are the measured attenuation values in two ROIs highlighted by circles in (a1) and (b1), which correspond to the iodine and muscle, respectively.

Physical experiment result

Tube phantom

Figure 8 shows the decomposition results of the water phantom. Images in the first row correspond to the decomposed CaCl $_{2}$ , iodine, oil, water basis images and the colour overlay image. Images in the second and third rows correspond to the FBP images reconstructed from the full scan projections of different tube voltages, and the CT images synthesized from the decomposed basis images. We can observe that four materials can be well separated from each other, and CT images of different kVs can be synthesized. Plots in Figure 8(d) and (e) compares the attenuation values of the FBP reconstructed and synthesized CT images. The quantitative results indicate that the SkV-Net could generate accurate multi-energy CT images with reduced noise levels. Figure 9 show the measured concentrations of calcium chloride and iodine. The linear fitting curves match the ground truth with minimum errors.

Figure 8.

Tube phantom results. Images from (a1) to (a5) are the decomposed CaCl $_{2}$ , iodine, oil, water images and the colour overlay image. The display windows for the four basis material images are [0, 300]mg/ml, [0, 20]mg/ml, [0, 1]g/ml, and [0.5, 1.2]g/ml, respectively. Images from (b1) to (b5) are FBP reconstructed CT images from the full scan projections. Images from (c1) to (c5) are CT images synthesized from the decomposed basis images. The display windows are [0, 0.4]cm⁻¹. The scale-bar denotes 35 mm. Plots in (d) and (e) are the measured attenuation values in two ROIs highlighted by circles in (b1) and (c1), which correspond to the iodine and CaCl $_{2}$ , respectively.

Figure 9.

Quantification results of (a) Iodine and (b) CaCl $_{2}$ in the tube phantom. Linear fittings (dashed lines) are performed.

Pork specimen

The decomposition results of the pork specimen is demonstrated in Figure 10. Images in the first row correspond to the decomposed bone, iodine, fat, muscle basis images and the colour overlay image. Images in the second and third rows correspond to the the FBP images reconstructed from the full scan projections of different tube voltages, and the CT images synthesized from the decomposed basis images. Plots in Figure 10(d) and (e) compares the attenuation values of the FBP reconstructed and synthesized CT images. It is observed that basis materials in the pork specimen can also be identified. Structural information are preserved, see the zoomed-in ROI in Figure 10(a1). The synthesized CT images at different kVs have close appearance and consistent attenuation values with reduced image noises compared to the FBP images, see Figure 10(d) and (e).

Figure 10.

Pork specimen results. Images from (a1) to (a5) are the decomposed bone, iodine, fat, muscle images and the colour overlay image. Display windows for four basis material images are [0, 1.4] g/cm $^{3}$ , [0, 20] mg/ml, [0, 1] g/cm $^{3}$ , and [0, 1.4] g/cm $^{3}$ , respectively. Images from (b1) to (b5) are FBP reconstructed CT images from the full scan projections of different tube voltages. Images from (c1) to (c5) are CT images synthesized from the decomposed basis images. The display windows are [0, 0.55]cm⁻¹. The scale-bar denotes 35 mm. Plots in (d) and (e) are the measured attenuation values in two ROIs highlighted by circles in (b1) and (c1), which correspond to the iodine and fat, respectively.

Computation time

In the current implementation, it takes approximately 0.14 seconds for the SkV-Net to decompose four $512 \times 512$ basis images from five $512 \times 512$ CT images. The computations are conducted on a workstation equipped with an Intel Xeon^® Silver 4210R CPU and an Nvidia RTX A6000 GPU. The particularly short computation time makes it promising to realize real-time material decomposition.

Discussions and conclusion

In this study, a dedicated multi-material decomposition network, SkV-Net, was developed for ultra-slow kV switching spectral CBCT. Compared to the fast kV switching technique, which alternates between low and high tube voltages in every other projection view, the ultra-slow kV switching technique requires only a few (five was assumed in this study) kV changes per gantry rotation. This makes it much easier to implement using existing CBCT systems. The proposed SkV-Net features a U-Net structure integrated with a multi-head axial attention mechanism. It takes the reconstructed CT images from different tube voltages as input and outputs multiple basis material density maps. Numerical simulations of the XCAT phantom, as well as physical experiments with a tube phantom and a pork specimen, were conducted. The results demonstrated that SkV-Net can accurately predict the density distributions of multiple basis materials.

While the preliminary results are promising, there are several limitations to the current work. First, due to the unavailability of ultra-slow kV switching hardware, only numerical simulations and experimentally synthesized data were used to evaluate the performance of SkV-Net. Further validation with real ultra-slow kV switching data is necessary once such CT systems become available. Second, the network design is still relatively empirical. More complex network architectures, such as transformer-based networks,³² generative adversarial networks,³³ conditional denoising diffusion probabilistic model (C-DDPM),³⁴ may provide additional performance improvements. Third, a more accurate imaging model is needed to generate high-quality numerical training data, accounting for factors such as detector responses, Compton scattering effects, and the X-ray source spectrum. Lastly, the tube voltages used in this study were limited to the range of 85 kV to 125 kV with 10 kV increments. Future studies could explore a broader range of acquisition settings with smaller increments.

In summary, a novel convolutional neural network, SkV-Net, has been developed for multi-material decomposition in ultra-slow kV switching spectral CBCT imaging. Both numerical and experimental results demonstrate the potential of this network for high-quality spectral CBCT imaging. It offers promising possibilities for the development of low-cost spectral CBCT systems in the future, which could enhance the capabilities of medical imaging in clinical practice.

Footnotes

ORCID iDs

Han Cui

Yongshuai Ge

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported in part by the Guangdong Basic and Applied Basic Research Foundation (2021TQ06Y108), the National Natural Science Foundation of China (62422123, 62201560, U23A20284).

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Fahrig

Jaffray

Sechopoulos

, et al. Flat-panel conebeam CT in the clinic: history and current state. J Med Imag 2021; 8: 052115.

Sauter

Muenzel

Dangelmaier

, et al. Dual-layer spectral computed tomography: Virtual non-contrast in comparison to true non-contrast images. Eur J Radiol 2018; 104: 108–114.

Pessis

Campagna

Sverzut

, et al. Virtual monochromatic spectral imaging with fast kilovoltage switching: reduction of metal artifacts at CT. Radiographics 2013; 33: 573–583.

Liu

Shapira

Chen

, et al. First-generation clinical dual-source photon-counting CT: ultra-low-dose quantitative spectral imaging. Eur Radiol 2022; 32: 8579–8587.

Wrazidlo

Walder

Estler

, et al. Radiation dose reduction in contrast-enhanced abdominal CT: comparison of photon-counting detector CT with 2nd generation dual-source dual-energy CT in an oncologic cohort. Acad Radiol 2023; 30: 855–862.

Petersilka

Bruder

Krauss

, et al. Technical principles of dual source CT. Eur J Radiol 2008; 68: 362–368.

Schmidt

Flohr

. Principles and applications of dual source CT. Phys Med 2020; 79: 36–46.

Rassouli

Etesami

Dhanantwari

, et al. Detector-based spectral CT with a novel dual-layer technology: principles and applications. Insights Imaging 2017; 8: 589–598.

Wang

Shapiro

, et al. Dual energy imaging with a dual-layer flat panel detector. In: Medical imaging 2019: Physics of medical imaging, Vol. 10948, pp.269–278. SPIE.

10.

Chandra

Langan

. Gemstone detector: dual energy imaging via fast kVp switching. In: Dual Energy CT in Clinical Practice. Springer, 2011, pp.35–41.

11.

Lee

. Rapid kV switching dual-energy CT imaging. In: Dual-Energy CT in Cardiovascular Imaging. Springer, 2015, pp.45–60.

12.

Petrongolo

Zhu

. Single-scan dual-energy CT using primary modulation. IEEE Trans Med Imaging 2018; 37: 1799–1808.

13.

Tivnan

Wang

Stayman

. A prototype spatial–spectral CT system for material decomposition with energy-integrating detectors. Med Phys 2021; 48: 6401–6411.

14.

Greffier

Van Ngoc Ty

Fitton

, et al. Spectral performance of two split-filter dual-energy CT systems: a phantom study. Med Phys 2023; 50: 6828–6835.

15.

Szczykutowicz

Chen

. Dual energy CT using slow kVp switching acquisition and prior image constrained compressed sensing. Phys Med Biol 2010; 55: 6411.

16.

Sisniega

Abascal

Abella

, et al. Iterative Dual-Energy material decomposition for slow kVp switching: a compressed sensing approach. In: XIII Mediterranean conference on medical and biological engineering and computing 2013: MEDICON 2013, 25-28 September 2013, Seville, Spain, pp.491–494. Springer.

17.

Zhu

, et al. Performance evaluation of quantitative material decomposition in slow kVp switching dual-energy CT. J Xray Sci Technol 2023; 32: 69–85.

18.

Szczykutowicz

Hsieh

Chen

. The dependence of image quality on the number of high and low kVp projections in dual energy CT using the prior image constrained compressed sensing (PICCS) algorithm. In: Medical imaging 2010: Physics of medical imaging, Vol. 7622, pp.678–687. SPIE.

19.

Chen

Zhang

Xia

, et al. Non-convex primal-dual algorithm for image reconstruction in spectral CT. Comput Med Imaging Graph 2021; 87: 101821.

20.

Sun

Yang

, et al. DIRECT-net: A unified mutual-domain material decomposition network for quantitative dual-energy CT imaging. Med Phys 2022; 49: 917–934.

21.

Abascal

Ducros

Pronina

, et al. Material decomposition in spectral CT using deep learning: a sim2Real transfer approach. IEEE Access 2021; 9: 25632–25647.

22.

Zhu

Zhang

, et al. Feasibility study of three-material decomposition in dual-energy cone-beam CT imaging with deep learning. Phys Med Biol 2022; 67: 145012.

23.

Cao

Shapira

Maidment

, et al. Hepatic dual-contrast CT imaging: slow triple kvp switching CT with CNN-based sinogram completion and material decomposition. J Med Imag 2022; 9: 014003.

24.

Cao

Shapira

Noël

. GAN-based sinogram completion for slow triple kVp switching CT. In: Medical Imaging 2021: Physics of medical imaging, Vol. 11595, 2021, pp.694–700. SPIE.

25.

Huang

Lin

Tong

, et al. Unet 3+: A full-scale connected unet for medical image segmentation. In: ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP), 2020, pp.1055–1059. IEEE.

26.

Yan

Tang

Sun

, et al. After-unet: Axial fusion transformer unet for medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp.3971–3981.

27.

Cao

Wang

Chen

, et al. Swin-unet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision, 2022, pp.205–218. Springer.

28.

Wang

Zhu

Green

, et al. Axial-deeplab: stand-alone axial-attention for panoptic segmentation. In: European conference on computer vision, 2020, pp.108–126. Springer.

29.

Deng

Dong

Socher

, et al. Imagenet: a large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition, 2009, pp.248–255. IEEE.

30.

Poludniowski

Landry

Deblois

, et al. SpekCalc: a program to calculate photon spectra from tungsten anode x-ray tubes. Phys Med Biol 2009; 54: N433.

31.

Hubbell

Seltzer

. Tables of X-ray mass attenuation coefficients and mass energy-absorption coefficients 1 keV to 20 MeV for elements Z

=

1 to 92 and 48 additional substances of dosimetric interest. Technical report, National Institute of Standards and Physics Laboratory, 1995.

32.

Zhang

Yan

, et al. TIME-net: Transformer-integrated multi-encoder network for limited-angle artifact removal in dual-energy CBCT. Med Image Anal 2023; 83: 102650.

33.

Peng

Chang

Xie

, et al. Image-domain material decomposition for dual-energy CT using unsupervised learning with data-fidelity loss. Med Phys 2024; 51: 6185–6195.

34.

Gao

Qiu

Xie

, et al. CT-based synthetic contrast-enhanced dual-energy CT generation using conditional denoising diffusion probabilistic model. Phys Med Biol 2024; 69: 165015.