1.

INTRODUCTION

X-raycomputed tomography (CT) is widely used in clinical diagnosis and treatment because of its ability to image the body’s three-dimensional anatomy in a non-invasive manner. However, ionizing radiation generated during CT scanning will accumulate in the human body, and high radiation doses will induce the risk of cancer in human tissues and organs.¹ Given these risks, efforts have been made on reducing the radiation dose and the principles of As Low As Reasonably Achievable (ALARA) is profoundly practiced in clinical CT imaging.² However, reducing the radiation dose inevitably increases the noise and artifacts of reconstructed CT images, which compromises the diagnostic performance. Consequently, improving the image quality of low-dose CT (LDCT) has become a hot topic in medical imaging over the past decade.

Recently, with the rapid development of deep learning (DL) technology, the LDCT imaging algorithm is dominated by convolutional neural network and has achieved unprecedented success. DL-based algorithm learns the mapping from LDCT projection/image to normal-dose CT (NDCT) ones by designing an elaborate convolutional neural network (CNN), such as RED-CNN,³ Wavelet networks,⁴ and Tensor-Net.⁵ A key factor in the success of these supervised algorithms is the availability of big data, that is, a large amount of paired LDCT and NDCT images.^{6, 7} Despite its superior denoising results, some issues still must be resolved before the DL models can be widely deployed in clinic. First, given the increase in total radiation dose, matched LDCT and NDCT data cannot be obtained in clinical practice. As a result, true NDCT and LDCT paired data are not available. Second, traditional LDCT simulation methods⁸ usually insert noise into the raw projection data, however, very few raw data are publicly available to simulate paired training data, which degrades the generalization performance of the network. Third, most of DL-based models are designed for specific dose levels but perform poorly at lower doses.

To solve above problems, in this work, we propose a dual-task learning network (DTNet) for low-dose CT simulation and denoising at arbitrary dose levels. The presented DTNet can integrate LDCT simulation and denoising into a unified framework by learning the joint distribution of LDCT and NDCT data. Specifically, we first propose to train the simulation network by learning a mapping from NDCT to LDCT at different levels, where the dose level can be continuously controlled by a noise factor. In the denoising task, we present a multi- level LDCT learning strategy, which uses LDCT data of different levels to train the denoising network. Once trained, DTNet can be used for multi-level LDCT images simulation and denoising simultaneously, which greatly improves the applicability and generalization performance of the model.

2.

METHODOLOGY

2.1

Low-dose CT Simulation Network

Let x ∈ 𝕀 ^H×Wdenotes a NDCT image and y ∈ 𝕀^H×W denotes the corresponding LDCT image. Low-dose CT simulation is an inverse process of LDCT denoising, which learns the opposite mapping from x to y. To precisely control the noise level of the generated LDCT image, we propose a noise control factor guided LDCT simulation scheme, as illustrated in Fig. 1 (a). Specifically, given a NDCT image and a mask image with value of 1, the simulator network encodes them into high-dimensional features for coupling and progressively decoding to reconstruct LDCT image of different levels, where the dose level can be continuously controlled by a noise factor. In addition, instead of directly applying the pixel-wise loss to the target image, we use a gaussian filter to extract the first-order statistics information of simulation and reference LDCT noise, and then constraint it with the MAE loss function, which can be formulated as:

Figure 1.

The overall structure of the proposed DTNet framework. (a) Low-Dose CT simulation module, (b) Low-Dose CT denoising module, (c) Dual-task learning module. m is the mask image with value of 1. k_j is the noise scale factor.

Here, m is the mask image with value of 1, k_j = 1, …, j is the noise factor that controls the simulated dose level j. θ_S represents the parameters of simulation network S. GF is a 2d Gaussian filter convolution kernel with a size of 5 x 5.

2.2

Low-dose CT Denoising Network

Typically, the DL-based denoising problem is to build a prediction network R(·) that learns the non-linear mapping from y to x, i.e., R : y → x. In clinical, CT images of various dose levels may be obtained to meet the clinical diagnosis demands. In Fig. 1 (b), we propose a multi-level low-dose CT learning strategy to train the denoising network with different levels low-dose CT data by minimizing the mean absolute error (MAE) loss function, which is expressed as:

Here,y_i ^j, i = 1, …, N, j = 1, …, M represents the LDCT images, where N is the total number of LDCT images and M is number of dose levels. θ_R represent the parameters of denoising network R. Once trained, the denoising model can be applied to LDCT reconstruction at different dose levels in the clinic.

2.3

Dual-task Learning Network

To jointly optimize low-dose CT simulation and denoising tasks, we adopt the dual-task learning network (DT- Net), which uses a joint discriminator to alternately optimize the simulation and denoising network by learning the joint distribution p(x, y^j) of LDCT and NDCT data, as shown in Fig. 1(c). Let p_S (x, y^j) and p_R (x, y^j) represent the pseudo joint distribution of simulation and denoising task, respectively. The dual adversarial loss can be defined as follows:

represent the pseudo joint distribution of simulation and denoising task, respectively. The dual adversarial loss can be defined as follows:

Here, E[·] denotes the expectation operator, D represents the discriminator, which is used to receive the image pair and distinguish them as real or fake samples. The hyper-parameters λ_R and λ_scontrols the weight of GAN loss.

The final objective function for optimizing DTNet can be formulated as:

where α and β are the weight parameters that control the trade-off between adversarial loss and fidelity loss of simulation and denoising tasks.

3.

EXPERIMENTS

4.

RESULTS

Fig. 2 shows the simulation and denoising results of abdominal CT image at five different dose levels: normal- dose, 1/4, 1/8, 1/16, 1/20-dose. In the simulation task, the corresponding simulation noise factors are set to 0, 1/4, 1/8, 1/16, and 1/20, respectively. Note that the noise factor equal to 0 means that it does not insert any noise into the NDCT image. It can be observed that the proposed method can simulate LDCT images of different dose levels and the learned noise intensity and characterization are similar to the reference images. In addition, we also calculate the noise power spectrum (NPS) maps of magnified ROIs to evaluate the statistical property of noise. We can observe that the NPS of generated LDCT images is very close to the reference images. When the noise factor is set to 0, only little noise is embedded in the output image, which can be seen from the ROI and the corresponding NPS. To verify the robustness of DTNet in simulating other dose levels which are not including in training data, as shown in Fig. 3. It can be seen that the noise intensity of simulated LDCT images continuously increases with the increase of noise factor k_j. This demonstrates that the proposed DTNet has the ability to simulate the realistic LDCT images and can control the noise level well.

Figure 2.

The simulation and denoising results at different dose levels of proposed DTNet. The display window of images and zoomed-in ROIs is [-140, 260] HU. The display window of NPS maps is [0 3000] HU²mm².

Figure 3.

Arbitrary dose levels simulation results of proposed DTNet. From the left to the right column, the noise level of CT image gradually increases as the noise factor k_j increases. The display window is [- 140, 260] HU

In the denoising task, the last row in Fig. 2 shows that DTNet can efficiently suppress noise and artifacts at difference dose levels. In particular, for normal-dose images with a small amount of noise, we can also remove the noise without smoothing the image content. And for ultra-low doses, such as 1/20-dose, some small structures are completely drowned out by noise and are difficult to recover well. Therefore, ultra-low dose scanning can be used for special imaging tasks where anatomical details are not important, such as localization imaging. To quantitatively analyze the denoising performance of DTNet, we calculate the peak-to-noise ratio (PSNR) and structural similarity (SSIM), as summarized in Table 1. We can see that DTNet obtains the best quantitative values at different dose levels compared to RED-CNN³ and WGAN-VGG,¹² which is consistent with the visual evaluation.

Table 1.

PSNR and SSIM Quantitative comparison of DTNet denoising results at different dose levels.

5.

CONCLUSION

In this paper, we have presented a dual-task learning network (DTNet) for LDCT simulation and denoising tasks. In the simulation task, the simulation network encodes the NDCT image and mask image into high-dimensional features for coupling and decoding to generate LDCT images at different dose levels, where the dose levels can be controlled by a noise factor. In the denoising task, the multi-level LDCT learning strategy is used to train the denoising network, which can learn many-to-one end-to-end mapping. The presented DTNet integrates the LDCT simulation and denoising tasks into a unified optimization model. Both the quantitative and qualitative evaluation results have demonstrated the promising performance of DTNet in terms of LDCT simulation and denoising. In the feature, we will further improve the performance of DTNet by incorporating advanced network and prior information.