cs.CVApr 8, 2026
SaveBalancing Efficiency and Restoration: Lightweight Mamba-Based Model for CT Metal Artifact Reduction
Organizations: New York University, America · BaiyunPort, China · University of Technology Sydney, Australia · Sanda University, China · Shan Dong Xiehe College, China · Tsinghua University, China · Gannan Normal University, China · Southern Medical University, China · Shenzhen Research Institute of Big Data, China
Abstract
In computed tomography imaging, metal implants frequently generate severe artifacts that compromise image quality and hinder diagnostic accuracy. There are three main challenges in the existing methods: the deterioration of organ and tissue structures, dependence on sinogram data, and an imbalance between resource use and restoration efficiency. Addressing these issues, we introduce MARMamba, which effectively eliminates artifacts caused by metals of different sizes while maintaining the integrity of the original anatomical structures of the image. Furthermore, this model only focuses on CT images affected by metal artifacts, thus negating the requirement for additional input data. The model is a streamlined UNet architecture, which incorporates multi-scale Mamba (MS-Mamba) as its core module. Within MS-Mamba, a flip mamba block captures comprehensive contextual information by analyzing images from multiple orientations. Subsequently, the average maximum feed-forward network integrates critical features with average features to suppress the artifacts. This combination allows MARMamba to reduce artifacts efficiently. The experimental results demonstrate that our model excels in reducing metal artifacts, offering distinct advantages over other models. It also strikes an optimal balance between computational demands, memory usage, and the number of parameters, highlighting its practical utility in the real world. The code of the presented model is available at: https://github.com/RICKand-MORTY/MARMamba.
Figures & tables
Fig. 1: MARMamba backbone architecture. It consists of upsampling and downsampling components, each comprising three stages, with MS-Mamba serving as the core module for every stage.
Fig. 2: Internal structure of MS-Mamba, which consists of two components: FMB and AMFN. The right side displays the internal details of those two modules.
| Model | Large | Medium | ||||||
| PSNR | SSIM | RMSE | LPIPS | PSNR | SSIM | RMSE | LPIPS | |
| InDuDoNet [ 1 ] | 39.77 ± 3.60 | 0.9881 ± 0.0051 | 3.30 ± 1.26 | 0.0576 ± 0.0176 | 41.72 ± 4.19 | 0.9926 ± 0.0022 | 2.76 ± 1.58 | 0.0430 ± 0.0111 |
| InDuDoNet+ [ 6 ] | 38.29 ± 3.36 | 0.9858 ± 0.0049 | 3.93 ± 1.97 | 0.0639 ± 0.0164 | 40.72 ± 3.74 | 0.9904 ± 0.0029 | 3.02 ± 1.54 | 0.0501 ± 0.0122 |
| OSCNet [ 12 ] | 41.14 ± 2.83 | 0.9910 ± 0.0023 | 2.77 ± 1.18 | 0.0476 ± 0.0128 | 41.08 ± 4.24 | 0.9923 ± 0.0023 | 3.00 ± 1.88 | 0.0401 ± 0.0112 |
| OSCNet+ [ 13 ] | 41.51 ± 2.96 | 0.9911 ± 0.0023 | 2.66 ± 1.17 | 0.0474 ± 0.0126 | 41.88 ± 3.88 | 0.9928 ± 0.0019 | 2.69 ± 1.61 | 0.0400 ± 0.0112 |
| MEPNet [ 7 ] | 38.28 ± 3.33 | 0.9852 ± 0.0073 | 3.90 ± 1.55 | 0.0648 ± 0.0194 | 42.90 ± 3.78 | 0.9922 ± 0.0021 | 2.36 ± 1.27 | 0.0488 ± 0.0144 |
TABLE I: Quantitative comparison experiments. Evaluate the effectiveness of metal artifact restoration for large and medium sizes using the PSNR( ), SSIM( ), RMSE( ), LPIPS( ) metrics (Non-metallic area) . Each grid cell represents the mean standard deviation. The optimal mean values highlighted in bold red .
| Model | Small | Tiny | ||||||
| PSNR | SSIM | RMSE | LPIPS | PSNR | SSIM | RMSE | LPIPS | |
| InDuDoNet [ 1 ] | 46.37 ± 2.17 | 0.9949 ± 0.0011 | 1.47 ± 0.41 | 0.0375 ± 0.0092 | 46.72 ± 2.13 | 0.9953 ± 0.0008 | 1.41 ± 0.40 | 0.0364 ± 0.0087 |
| InDuDoNet+ [ 6 ] | 46.72 ± 2.10 | 0.9940 ± 0.0013 | 1.41 ± 0.38 | 0.0414 ± 0.0101 | 47.42 ± 1.80 | 0.9946 ± 0.0009 | 1.29 ± 0.29 | 0.0392 ± 0.0096 |
| OSCNet [ 12 ] | 46.24 ± 2.46 | 0.9948 ± 0.0012 | 1.51 ± 0.46 | 0.0339 ± 0.0090 | 47.17 ± 2.26 | 0.9952 ± 0.0009 | 1.35 ± 0.44 | 0.0325 ± 0.0091 |
| OSCNet+ [ 13 ] | 46.49 ± 2.45 | 0.9948 ± 0.0012 | 1.47 ± 0.44 | 0.0339 ± 0.0093 | 47.37 ± 2.33 | 0.9952 ± 0.0009 | 1.32 ± 0.43 | 0.0325 ± 0.0092 |
| MEPNet [ 7 ] | 47.79 ± 1.79 | 0.9943 ± 0.0012 | 1.24 ± 0.28 | 0.0430 ± 0.0122 | 48.16 ± 1.70 | 0.9946 ± 0.0011 | 1.19 ± 0.28 | 0.0413 ± 0.0120 |
TABLE II: Quantitative comparison experiments. Evaluate the effectiveness of metal artifact restoration for small and tiny sizes using the PSNR( ), SSIM( ), RMSE( ), LPIPS( ) metrics (Non-metallic area) . Each grid cell represents the mean standard deviation. The optimal mean values highlighted in bold red .
| Model | Large | Medium | ||||||
| PSNR | SSIM | RMSE | LPIPS | PSNR | SSIM | RMSE | LPIPS | |
| InDuDoNet [ 1 ] | 34.62 ± 2.95 | 0.9801 ± 0.0074 | 5.83 ± 1.93 | 0.0708 ± 0.0214 | 37.54 ± 3.47 | 0.9893 ± 0.0036 | 4.31 ± 2.21 | 0.0485 ± 0.0126 |
| InDuDoNet+ [ 6 ] | 34.40 ± 4.02 | 0.9792 ± 0.0069 | 6.29 ± 2.96 | 0.0768 ± 0.0187 | 39.47 ± 3.61 | 0.9884 ± 0.0035 | 3.45 ± 1.60 | 0.0550 ± 0.0129 |
| OSCNet [ 12 ] | 39.20 ± 3.00 | 0.9874 ± 0.0035 | 3.47 ± 1.37 | 0.0582 ± 0.0156 | 40.28 ± 4.05 | 0.9906 ± 0.0030 | 3.24 ± 1.86 | 0.0443 ± 0.0122 |
| OSCNet+ [ 13 ] | 39.60 ± 3.06 | 0.9876 ± 0.0035 | 3.32 ± 1.36 | 0.0578 ± 0.0152 | 41.09 ± 3.77 | 0.9912 ± 0.0025 | 2.91 ± 1.63 | 0.0442 ± 0.0119 |
| MEPNet [ 7 ] | 36.53 ± 3.07 | 0.9795 ± 0.0090 | 4.71 ± 1.70 | 0.0764 ± 0.0228 | 40.68 ± 3.45 | 0.9900 ± 0.0028 | 2.98 ± 1.34 | 0.0532 ± 0.0152 |
TABLE III: Quantitative comparison experiments. Evaluate the effectiveness of metal artifact restoration for large and medium sizes using the PSNR( ), SSIM( ), RMSE( ), LPIPS( ) metrics (metal-included region) . Each grid cell represents the mean standard deviation. The optimal mean values highlighted in bold red font .
| Model | Small | Tiny | ||||||
| PSNR | SSIM | RMSE | LPIPS | PSNR | SSIM | RMSE | LPIPS | |
| InDuDoNet [ 1 ] | 43.81 ± 3.98 | 0.9937 ± 0.0022 | 2.25 ± 2.11 | 0.0410 ± 0.0101 | 46.57 ± 2.11 | 0.9949 ± 0.0009 | 1.44 ± 0.40 | 0.0383 ± 0.0091 |
| InDuDoNet+ [ 6 ] | 45.95 ± 2.52 | 0.9935 ± 0.0015 | 1.56 ± 0.49 | 0.0446 ± 0.0105 | 46.98 ± 2.01 | 0.9942 ± 0.0011 | 1.37 ± 0.33 | 0.0416 ± 0.0098 |
| OSCNet [ 12 ] | 46.02 ± 2.49 | 0.9944 ± 0.0013 | 1.55 ± 0.47 | 0.0364 ± 0.0097 | 47.10 ± 2.27 | 0.9950 ± 0.0009 | 1.36 ± 0.44 | 0.0343 ± 0.0096 |
| OSCNet+ [ 13 ] | 46.26 ± 2.49 | 0.9945 ± 0.0013 | 1.51 ± 0.46 | 0.0364 ± 0.0099 | 47.30 ± 2.34 | 0.9950 ± 0.0009 | 1.33 ± 0.43 | 0.0342 ± 0.0097 |
| MEPNet [ 7 ] | 45.80 ± 1.67 | 0.9934 ± 0.0013 | 1.55 ± 0.34 | 0.0461 ± 0.0131 | 47.98 ± 1.66 | 0.9943 ± 0.0011 | 1.21 ± 0.28 | 0.0433 ± 0.0126 |
TABLE IV: Quantitative comparison experiments. Evaluate the effectiveness of metal artifact restoration for small and tiny sizes using the PSNR( ), SSIM( ), RMSE( ), LPIPS( ) metrics (metal-included region) . Each grid cell represents the mean standard deviation. The optimal mean values highlighted in bold red font .
| Model | Time/s | Param/M |
|---|---|---|
| InDuDoNet [ 1 ] | 0.2537±0.0037 | 5.09 |
| InDuDoNet+ [ 6 ] | 0.2521±0.0041 | 1.75 |
| OSCNet [ 12 ] | 0.0181±0.0004 | 1.27 |
| OSCNet+ [ 13 ] | 0.0243±0.0006 | 1.28 |
| MEPNet [ 7 ] | 0.3585±0.0168 | 4.72 |
| DICDNet [ 41 ] | 0.0187±0.0007 | 1.28 |
TABLE V: Parameters vs. Inference Time
Fig. 3: Large-scale visual comparison of metallic implants. The input image is selected from the SynDeepLesion [ 1 ] dataset. The red regions in the images indicate the locations of metallic components. To better visualize fine details, only the content within the blue bounding boxes is displayed for each model. A Hounsfield Unit (HU) window of [-175 HU, 275 HU] is used for visualization.
Fig. 4: Medium metallic implants. The input image is selected from the SynDeepLesion [ 1 ] dataset. The HU window is set to the range of [-175 HU, 275 HU].
Fig. 5: Small metallic implants. The input image is selected from the SynDeepLesion [ 1 ] dataset. The HU window is set to the range of [-175 HU, 275 HU].
Fig. 6: Tiny metallic implants. The input image is selected from the SynDeepLesion [ 1 ] dataset. The HU window is set to the range of [-175 HU, 275 HU].
| Model | Score |
|---|---|
| InDuDoNet [ 1 ] | 3.489±1.223 |
| InDuDoNet+ [ 6 ] | 3.586±1.194 |
| OSCNet [ 12 ] | 3.409±1.252 |
| OSCNet+ [ 13 ] | 3.863 ±1.120 |
| ACDNet [ 11 ] | 3.806±1.217 |
| MEPNet [ 7 ] | 3.294±1.310 |
TABLE VI: Real‑World Scene Restoration Effectiveness Rating (1–5 Scale)
Fig. 7: Visual comparison in real-world scenarios. Excluding the input image, the visual images of the other models display the inner regions marked by the red rectangles. The display window is [-175HU, 275HU].
Fig. 8: Comparison of computational complexity and memory overhead. Bubble size is proportional to . Disable gradient during evaluation.
Fig. 9: Image size vs. resource consumption. Disable gradient during testing.
| Normal | Horizontal | Vertical | PSNR/SSIM/RMSE | Param/M | Time/ms |
|---|---|---|---|---|---|
| ✓ | 48.2703/0.9938/1.2469 | 0.47 | 32.4 | ||
| ✓ | ✓ | 48.6277/0.9941/1.1908 | 0.53 | 34.1 | |
| ✓ | ✓ | ✓ | 49.2559/0.9948/1.0784 | 0.59 | 36.2 |
TABLE VII: Ablation experiments within the FMB. ✓indicating enablement of each branch in the orientation scanning stage. Red bold font indicates the best value.
Fig. 10: Feature visualization of the FMB. From left to right, the panels illustrate the input image, the Mamba output corresponding to the original (unflipped) input, the output of the Mamba block under horizontal flipping, and the output under vertical flipping. The regions enclosed by red rectangles indicate the locations of metallic implants, which serve as the primary sources of artifacts. The display window of input image is [-175HU, 275HU].
| Average | Maximum | PSNR/SSIM/RMSE | Time/ms | |
|---|---|---|---|---|
| ✓ | 48.5776/0.9940/1.2013 | 36.6 | ||
| ✓ | 48.4413/0.9940/1.2337 | 35.3 | ||
| ✓ | ✓ | 49.2559/0.9948/1.0784 | 36.2 |
TABLE VIII: Ablation experiments within the AMB. ✓indicating enablement of each pooling layer. Red bold text highlights the optimal value.
Fig. 11: Directional energy log-ratio of MS-Mamba branches. Positive values indicate orientations where the flipped branches selectively enhance structural responses relative to the non-flipped branch, while negative values indicate suppression.
| Loss | PSNR /SSIM /RMSE |
|---|---|
| Pseudo-Huber | 48.4497/0.9944/1.1849 |
| LPIPS | 46.1632/0.9931/1.5145 |
| Pseudo-Huber+LPIPS | 49.2559/0.9948/1.0784 |
TABLE IX: Loss Function Ablation Experiment.
Fig. 12: Error maps between restored images and ground truths. Samples selected from the SynDeepLesion test set [ 1 ] . Magnified details are provided below each image for observation. Red indicates the absolute error between the repaired image and the ground truth. To highlight the errors, the error values have been amplified by a factor of two. The metal-containing area was preserved in the restored image. HU window: [-175 HU, +275 HU].