SinoSynth: A Physics-based Domain Randomization Approach for Generalizable CBCT Image Enhancement

Our CBCT simulation is performed in the sinogram domain. The CT slice $x$ represented in Hounsfield Units is first converted into the linear attenuation coefficient map $x_{\mu}$ via a linear transformation [5]. Radon transform $\mathcal{R}$ integrated along the cone beam geometry $G$ is then applied to project $x_{\mu}$ onto the sinogram:

where the constant $I_{0}(E_{i})$ is the intensity of the entry X-ray beam within composite energy levels $\{E_{i}|i\in[1,N]\}\subseteq[20,120]$ keV. We then simulate different types of CBCT artifacts and noise based on the derived sinogram $x_{s}$. Reconstructing from the sinogram using filtered back projection gives the corresponding CBCT image that is used for training. We implemented Eq 1 with the ASTRA Toolbox [26] and Operator Discretization Library ²²2https://github.com/odlgroup/odl.

Scanner Effects Simulation. The diversity among different CBCT scanners significantly impacts the varied appearance of CBCT artifacts, resulting in decreased performance of the network [18]. To simulate the scanner variability, we parameterize CBCT scanners with different cone beam geometries $G(d_{1},d_{2},n,s)$. As illustrated in Fig. 2 (b), $d_{1}\in[10,100]$ denotes the distance from the X-ray beam source to the patient; $d_{2}\in[10,100]$ is the distance from the patient to the detector; $n\in[64,512]$ denotes the number of X-ray projections, and a smaller $n$ indicates a sparser view; $s\in[128\times 128,512\times 512]$ denotes the detector size. Hence, we simulate the effects of CBCT scanners from different vendors by computing Eq. 1 with randomly parameterized $G$.

Metal Artifacts usually occur in the presence of metal implants. The low X-ray transmission of metallic implants and the polychromatic nature of the X-ray source result in severe beam hardening [2], bringing scatter and streak artifacts [30]. Thus, for simulation, we need to create metal trace regions on the sinogram $x_{s}$. This is done by first taking the intersection between a random mask $m_{t}$ obtained through cubic bézier curve [12] controlled by randomly sampled points, and a mask of bone areas $m_{b}$ obtained through thresholding, $m_{\Omega}=m_{t}\cap m_{b}$, and converting $m_{\Omega}$ to the sinogram $m_{\Omega}^{s}$. Then, we update the metal trace regions of $x_{s}$ by filling in the linear attenuation coefficients of the metal implanted area:

where each $metal_{\mu}(E_{i})$ corresponding to an energy level $E_{i}$ is computed as: $metal_{\mu}(E_{i})=m_{m}\cdot\mu_{m}(E_{i})\cdot(\rho_{m}-1.0)$. The mass attenuation coefficient $\mu_{m}(E_{i})$ can be obtained from [14]. Adjusting the density of the metal material $\rho_{m}>1$ allows for controlling the intensity of the metal artifacts.

Extinction Artifacts occur when the object contains highly absorbent material, which significantly attenuates the X-ray signal, reducing it to near-zero [23]. Since the attenuated regions on CBCT images are often irregular and discontinuous, we generate a random image-domain mask $m_{et}^{s}$ using the cubic bézier curve controlled by randomly sampled points within a local region. The sinogram $x_{s}$ is then updated by:

where $\mathcal{R}$ denotes Radon transform, $\lambda_{et}\in[0,1]$ controls the attenuation extent.

Quantum Noise is the main source of image deterioration in plain radiography, which arises from the natural variability in how photons reach the detector [28]. As quantum noise predominantly impacts images acquired using low radiation doses, CBCT images are expected to exhibit a higher noise level compared to CT images. Given that quantum noise usually follows a Poisson distribution [24], the quantum noise on sinogram $x_{s}$ can be simulated as:

where $\alpha$ controls the noise level, and $k$ is the number of photon occurrences during a time interval.

Zebra artifacts appear as alternating bright and dark stripes in CBCT images due to helical interpolation [21]. We simulate them by creating a binary mask $m_{z}$ with $n\sim\mathcal{U}(0,N)$ stripes. The width of each stripe is randomized within the range of $[1,10]$ pixels. We then simulate stripes of different directions by applying rotation to the mask $m_{z}=\texttt{Rotate}(m_{z},\theta)$, $\theta\in[0,\pi)$. The zebra artifact on the sinogram $x_{s}$ is simulated as:

Motion Artifacts are caused by the motion of the patient that appears as unsharpness in the reconstructed image [21]. The process can be viewed as an image $x$ disturbed by small deformations and displacements $\mathcal{D}$ during the imaging process, along with a blurring effect simulated by the Gaussian filter:

where $\mathcal{G}$ is the Gaussian blur filter. $\mathcal{DM}(\sigma_{d},r,z)$ represents the deformation and displacement caused by the motion, controlled by the deformation degree $\sigma_{D}\in[0.5,1.5]$, the rotation angle $r\in[0,\frac{\pi}{60}]$ and the resizing factor $z\in[0.9,1.1]$. $p$ is a factor balancing $x$ and its deformed result $\mathcal{D}(\cdot)(x)$.

As our CBCT simulation model $\mathcal{D}$ is differentiable, we propose simple anatomical structure constraints to regularize the generation of sCT in both the sinogram and image domains. For a simulated CBCT image $y$ and the network output $G_{\theta}(y)$, the denoised output should be converted back to $y$ via the degradation model. Meanwhile, the output should be consistent with the reference CT in the sinogram domain after being applied with a mask $M$ that removes the metal-affected region. Hence, the generative network $G_{\theta}$ is guided by the following losses to generate anatomically consistent output:

Since we applied our method to existing frameworks, all models were trained with their original losses (e.g., GAN losses) alongside the proposed structure-preserving losses.

We curated a Head and Neck CBCT dataset with 250 patients from five hospitals in Europe and the US with mAs used in scanning (10, 12, 20, and 50). 160 randomly sampled patient data are used for training, 20 for validation, and 70 for evaluation. All the data are resampled to a voxel size of $1\times 1\times 2\textit{mm}^{3}$, with an image size of $256\times 256$ for each slice. We utilize Peak-to-Signal Ratio (PSNR [13]), Structure Similarity (SSIM [13]), and Mean Absolute Error (MAE) of the HU values for evaluating all compared methods. Our method simulates CBCT images on the fly during training instead of pre-generating them. Each model is trained for 200 epochs.

We compare the established image-to-image translation models trained on our simulated data to the same models trained on the actual CBCT-CT data. The base models include supervised U-Net [6], ROI-aware DCLGAN [8], DRIT [16], and FGDM [17] based on diffusion probabilistic model [29]. Table 3 and Fig. 4 demonstrate that our method significantly enhances generalization across various artifacts, leading to the superior performance of CBCT noise reduction and soft tissue enhancement. Our method also improves the inpainting performance of the shoulder region.

We compare SinoSynth with existing augmentation methods [9, 4]. CycleGAN [7] is employed as the base model. As shown in Fig. 3, and Table 4, SinoSynth outperforms the existing augmentation methods. This is because the existing methods inadequately amplify the diversity of CBCT-specific artifacts, rendering the network susceptible to out-of-distribution CBCT artifacts, as depicted in Fig. 1. Our approach addresses this limitation by explicitly simulating a wide range of CBCT artifacts, thereby enhancing the performance of the CycleGAN.

Quantitative results in Table. 3 reveal two key insights. Firstly, for ablating the artifact types, PSNR/SSIM scores imply the occurrence frequency of the CBCT artifacts present in the test data. Secondly, for ablating the loss function design choices, both constraints contribute to CBCT image enhancement. As shown qualitatively in the appendix, the structure consistency constraint plays a crucial role in synthesizing structurally consistent images. The sinogram consistency constraint aids both CT noise reduction and accurate structure reconstruction.