MAR-DTN: Metal Artifact Reduction using Domain Transformation Network for Radiotherapy Planning

TL;DR

This paper presents MAR-DTN, a UNet-inspired domain transformation network that converts artifact-contaminated kVCT to artifact-free MVCT for head-and-neck radiotherapy planning. By aligning kVCT and MVCT slices, applying targeted preprocessing, and training with carefully chosen loss functions (notably $\mathcal{L}_1^{\mathbf{w}}$ and $\mathcal{L}_{\textit{FFL}}^{\beta,\alpha}$), MAR-DTN achieves state-of-the-art PSNR and SSIM on artifact regions and across whole volumes, outperforming pix2pix, custom-pix2pix, SwinIR, and INet in various configurations. Clinical feedback corroborates the quantitative gains, noting improved soft tissue and bone contrast in synthetic MVCTs, which can enhance radiotherapy planning while potentially reducing the need for additional imaging. The work demonstrates the practicality of a lightweight, domain-translation approach to metal artifact reduction in CT, with plans to expand datasets and generalize to broader anatomical regions and whole-body MAR tasks.

Abstract

For the planning of radiotherapy treatments for head and neck cancers, Computed Tomography (CT) scans of the patients are typically employed. However, in patients with head and neck cancer, the quality of standard CT scans generated using kilo-Voltage (kVCT) tube potentials is severely degraded by streak artifacts occurring in the presence of metallic implants such as dental fillings. Some radiotherapy devices offer the possibility of acquiring Mega-Voltage CT (MVCT) for daily patient setup verification, due to the higher energy of X-rays used, MVCT scans are almost entirely free from artifacts making them more suitable for radiotherapy treatment planning. In this study, we leverage the advantages of kVCT scans with those of MVCT scans (artifact-free). We propose a deep learning-based approach capable of generating artifact-free MVCT images from acquired kVCT images. The outcome offers the benefits of artifact-free MVCT images with enhanced soft tissue contrast, harnessing valuable information obtained through kVCT technology for precise therapy calibration. Our proposed method employs UNet-inspired model, and is compared with adversarial learning and transformer networks. This first and unique approach achieves remarkable success, with PSNR of 30.02 dB across the entire patient volume and 27.47 dB in artifact-affected regions exclusively. It is worth noting that the PSNR calculation excludes the background, concentrating solely on the region of interest.

Figures (5)

• Figure 1: (a) Abstract overview of proposed Domain Transformation Network. (b) Sagittal view of the body with distinct delineations of the head, neck, and body regions(blue). (c) kVCT (top) and MVCT (bottom) axial artifact slices after normalization and masking.
• Figure 2: Steps followed for dataset generation. We start with raw and unaligned kVCT and MVCT volumes --slices (lines in the cube) do not correspond. Then, volumes are pixel-aligned and so the slices correspond (Section \ref{['sub:alignmentPreprocessing']}). Finally, corresponding slices in kVCT and MVCT volumes are normalized and masked (Section \ref{['sub:alignmentPreprocessing']}).
• Figure 3: PSNR (a) and SSIM (b) values evaluated on the $\mathcal{D}_\text{All}$. The dots represent the mean value of all slices in the dataset, while the bars represent the mean value of slices with artifacts. Values obtained using the four considered networks (MAR-DTN, pix$2$pix, custom-pix$2$pix and SwinIR) trained on the $\mathcal{D}_\text{All}$ with the $\mathcal{L}_1^{w}$ loss function only.
• Figure 4: Heatmaps with the mean values of PSNR (a) and SSIM (b) evaluated on the test dataset after training the networks using the $\mathcal{L}_{\textit{FFL}}^{\beta,\alpha}$ loss function with various combinations of the parameters $\alpha$ and $\beta$ (x and y-axis, respectively). Each cell represents the mean of $8$ values, the first $4$ corresponding to the parameter value evaluated on $\mathcal{D}_\text{Art}^{Ts}$, and the last $4$ corresponding to the parameter value evaluated on the $\mathcal{D}_\text{All}^{Ts}$, for each neural network in the study, MAR-DTN, pix$2$pix, custom-pix2pix, and SwinIR, respectively.
• Figure 5: Reconstruction of a slice with artifacts by the different models and loss functions. First row shows preprocessed kVCT and MVCT images (ground truth). First column indicates the loss function, and the following ones indicate the model used. Networks have been trained on the $\mathcal{D}_\text{Art}$.