2.1

Pre-trained Clinical Module(PCM)

PCM allows the reconstruction process to converge to a clinically clean CT. The training process of the PCM module involves only clinically clean CT and the parameters of the module are kept frozen in the subsequent stages. As shown in Fig.2(a), clinical clean CT data I_c ∈ ℝ^H×W×1 are first embedded by an encoder E to obtain their corresponding compression features x_c ∈ ℝ^m×n×1. According to the VQGAN,¹² each pixel point of x_c ∈ ℝ^m×n×1 is nearest-neighbour computed with the feature item c_k ∈ ℝ^d in the codebook C to obtain its code sequence S ∈ {0,1,…,n−1}^m×n. Based on the sequence S, the features x_c are re-extracted from the codebook to form a new quantized feature X_cb ∈ ℝ^m×n×d, i.e., when S^i×n+j = k. The decoder reconstructs this clean CT data using X_cb as input. We adopt three image-level loss between the input I_c and the reconstruction result I_rec: L1 loss L₁, perceptual loss L_per and adversarial loss L_adv. In order to better optimise the codebook, we also adopt the code-level loss L_code:

where sg is the stop-gradient operator like VQGAN,¹² and β = 0.25 is a weight trade-off. The complete optimisation objective for this stage are as follows:

where α_l1 and α_adv is set to 1.5 and 0.3, respectively. After adequate training, the frozen codebook and decoder comprise the PCM to participate in the reconstruction process.

2.2

Robust Transformation Module(RTM)

RTM makes intermediate feature conversion more robust. The nearest neighbour algorithm would lead to mismatching of features due to metal artifacts on CT image pixels. We refer to the idea of CodeFormer¹³ to replace the nearest-neighbor algorithm of the VQGAN with the RTM module to match features in a more robust way.

Based on the pre-trained VQGAN in Sec.2.1, as shown in Fig.2(b), we accessed a Transformer-encoder¹⁴ E_tr containing nine self-attention blocks after the encoder E and added an extra linear layer for generating code sequences to form the RTM, as shown in Fig.3(b). Specifically, a CT image containing metal artifact is processed by encoder E_m (finetuned from E) to generate the corresponding artifact-affected feature x_m. The RTM takes xm as input and converts get the code sequence S^m(like S in Sec.2.1). The frozen PCM takes S^m as input to generate the quantized clean CT feature X_mcb and reconstructed clean CT image. We only use the code-level loss to train RTM: cross-entropy loss and L2 loss :

Figure 3.

The structure and calculation process of FGM(a) and RTM(b).

where the ground truth of code sequence S and feature X_cb is obtained from the pre-trained VQGAN in Sec.2.1. While training the RTM, E will be fine-tuned to get E_m, the complete loss of this stage is as follows:

where λ is set to 0.5.

2.3

Fidelity Guarantee Module(FGM)

FGM ensures the fidelity of the result by fusing the encoder stage features with the decoder features in a scaled manner, as shown in Fig.2(c). RTM replaces artifact-affected features directly by matching them with the codebook entries, and the whole process is discrete and a brand new feature map is obtained, so it is difficult to ensure the fidelity of the final result. Therefore, we make a new design according to the actual situation of metal artifact reduction, as shown in Fig.3(a). Specifically, we feed the intermediate stage features of the encoder to two convolution stacks f₁, f₂ with the same structure to obtain x_f1, x_f2. According to cGAN,¹⁵ we reshape the corresponding stage features of the decoder by α and β:

We add the FGM at the stage where the encoder and decoder generate features of size s ∈ {32, 64, 128, 256}, respectively. To ensure that enough coded feature information is involved in the reconstruction process and no additional noise is introduced, we set ω to 0.5. To train the FGM and fine-tune the E_m and RTM in Sec.2.2, we continue to use the metal-affected data as input and combine the loss L_PCM and L_RTM as the complete loss of this stage. The percentage of each loss function remains constant.

3.1

Experimental configuration

3.2

Experimental results

To validate the effectiveness of DCCGN in the field of image-only MAR, we compared it with one conventional MAR methods¹⁶ and several representative state-of-the-art (SOTA) image-only methods^{3, 9–11} on synthetic and clinical datasets, respectively.

Comparison with SOTAs on synthetic datasets.

As illustrated in Tab.1, the DCCGN outperformed the state-of-the-art (SOTA) methods with a high PSNR and SSIM scores for image-domain-only MAR in synthetic datasets. Qualitative comparisons, as shown in Fig.4, indicate that the DCCGN is more effective than other methods in reproducing the actual appearance and structural details and in circumventing the problem of noise residuals.

Figure 4.

Qualitative results demonstrate that DCCGN’s results are closer to artifact-free images without secondary artifacts. Red area indicates the metal implant; border color indicates the correspondence between the magnified view and the overview view.

Figure 5.

Comparison on clinical datasets.

Table 1

Quantitative comparisons (PSNR/SSIM) demonstrate the superior performance of DCCGN compared to SOTA image-domain-only methods.

Method	Domain-usage	PSNR↑	SSIM↑
Linear16	Conventional	26.01	0.8741
CNNMAR10	Image-only	30.38	0.9644
cGAN3	Image-only	34.10	0.9341
DICDNet11	Image-only	37.86	0.9788
OSCNet9	Image-only	42.19	0.9931
DCCGN	Image-only	41.83	0.9986