Organizations: Institute of Science and Technology for Brain-inspired Intelligence, Fudan University, Shanghai 200433, China · School of Instrumentation Science and Engineering, Harbin Institute of Technology, Harbin 150001, China · School of Information Engineering, Nanchang University, Nanchang 330031, China · School of Cyber Science and Engineering, Sichuan University, Chengdu 610065, China
Computed tomography (CT) restoration methods often use a single network for low-dose CT (LDCT), sparse-view CT (SVCT), and metal artifact reduction (MAR), despite the distinct physics of these degradations. We propose AgentCT, a vision-language agent that identifies the degradation and routes images to task-specific restoration models. A LoRA-adapted Qwen3-VL-4B controller predicts degradation type and severity, selects a specialist, and reassesses the restored image. The LDCT specialist uses an image-domain diffusion model, whereas the SVCT and MAR specialists combine projection-domain correction, data consistency, differentiable filtered back-projection, and image refinement. We evaluated the controller and specialists separately on patient-disjoint Mayo 2020 abdominal CT data across three severity levels for each degradation. The controller achieved 95.7% accuracy for degradation type and 93.9% for severity, with a 100% JSON parse rate. Using ground-truth degradation labels for routing, AgentCT achieved an overall PSNR/SSIM of 46.18 dB/0.984, compared with 44.93 dB/0.976 for AMIR and 44.29 dB/0.979 for TAMP, and attained the highest PSNR in all nine settings. For SVCT, its dual-domain specialist achieved 46.04 dB with 4.2 million parameters, compared with 44.43 dB for single-task AMIR and 43.57 dB for all-in-one AMIR with 23.5 million parameters. These results support degradation-specific CT restoration with controller-based model selection; evaluation under automatic routing and on independent clinical data remains necessary. Code is available at https://github.com/hao1635/AgentCT.
Figures & tables
Figure 1: Overview of AgentCT. A LoRA-adapted Qwen3-VL-4B controller identifies the degradation type and severity and selects a restoration specialist. LDCT uses an image-domain diffusion denoiser; SVCT and MAR use dual-domain networks. The restored image is returned to the controller for assessment, with at most one additional restoration call. Inputs classified as clean bypass restoration.
Method
Type accuracy
Severity accuracy
JSON parse rate
AgentCT controller
95.7%
93.9%
100%
Table 1: Perception performance of the vision-language controller on the patient-disjoint test set.
Category
Degraded
AMIR
TAMP
AgentCT ( ours )
Δ
LDCT 1/2
35.84 / .778
47.19 / .981
47.51 / .989
48.17 / .989
+0.66
LDCT 1/4
33.16 / .651
46.26 / .978
46.41 / .986
47.04 / .987
+0.63
LDCT 1/10
28.91 / .441
44.70 / .972
44.51 / .980
44.97 / .980
+0.27
SVCT 120v
37.45 / .855
48.49 / .989
47.64 / .989
50.35 / .993
+1.86
SVCT 60v
30.78 / .635
43.56 / .965
43.41 / .973
46.23 / .985
+2.67
SVCT 30v
25.50 / .468
38.65 / .943
38.90 / .945
41.53 / .967
+2.63
Table 2: Quantitative comparison on the nine CT degradation categories (PSNR/SSIM), evaluated in the [−1000,2000] HU window over the full patient-disjoint test set. Best PSNR/SSIM in bold; Δ is the PSNR gain of AgentCT over the second-best method in each row.
Category
AMIR
TAMP
AgentCT ( ours )
LDCT 1/2
14.4 ± 1.3
13.2 ± 0.9
12.8 ± 1.4
LDCT 1/4
15.8 ± 1.4
14.9 ± 1.2
14.4 ± 1.5
LDCT 1/10
18.6 ± 1.7
18.4 ± 2.1
17.9 ± 2.2
SVCT 120v
12.4 ± 1.5
13.0 ± 1.2
10.4 ± 1.9
SVCT 60v
20.9 ± 2.0
20.8 ± 1.9
15.6 ± 2.4
SVCT 30v
36.0 ± 5.8
34.6 ± 3.6
26.0 ± 4.2
Table 3: Restoration error (RMSE in Hounsfield units), mean ± standard deviation over the full patient-disjoint test set. Best in bold.
Figure 2: Qualitative comparison on LDCT, SVCT, and MAR. Columns show the degraded image, AMIR, TAMP, AgentCT, and the clean reference. The region of interest is magnified for visual comparison, and PSNR/SSIM are reported in the [−1000,2000] HU window.
Figure 3: Absolute error maps ∣x^−x∣ in Hounsfield units for the degraded input, AMIR, TAMP, and AgentCT on the three degradations (rows: LDCT, SVCT, MAR), sharing the color scale on the right. Darker indicates smaller error.
Figure 4: CT-value profiles in Hounsfield units along the yellow line segments marked on the reference images (rows: LDCT, SVCT, MAR). Curves represent the degraded input (gray), AMIR (blue), TAMP (green), AgentCT (red), and full-dose reference (black).
Figure 5: End-to-end restoration with AgentCT. The controller perceives the degradation, generates a structured routing action, invokes the corresponding specialist, and re-perceives the restored result for verification.
Method
SVCT overall PSNR / SSIM
AMIR single-task (SVCT only)
44.43 / .973
AMIR all-in-one ({LDCT, SVCT, MAR})
43.57 / .966
AgentCT (dual-domain, 4.2 M)
46.04 / .982
Table 4: Effect of physics-aware specialization on SVCT (overall PSNR/SSIM over the three sparsity levels).
Computed tomography uses X-ray measurements from multiple angles to reconstruct cross-sectional and volumetric images for diagnosis and scientific imaging.
Tsinghua University, Beijing, China · Data Science & Artificial Intelligence Research Institute, China Unicom, Beijing, China · Hunan University, Changsha, China +2