One network no longer handles all three CT degradations: AgentCT, a vision-language agent, routes each scan to a specialist and leads PSNR across nine settings

CT restoration shifts from one network handling every degradation to judging the degradation first and calling the matching specialist, so low-dose, sparse-view and metal-artifact cases each follow a path suited to their physics. Earlier universal restoration models such as AMIR and TAMP covered multiple degradations with a shared architecture, requiring a human to pick a model per scanning condition or letting a single network handle all of them. On the patient-disjoint Mayo 2020 abdominal CT test set, routing with ground-truth degradation labels, AgentCT reached an overall 46.18 dB/0.984, against 44.93 dB/0.976 for AMIR and 44.29 dB/0.979 for TAMP, and attained the highest PSNR in all nine settings.

The controller turns one windowed CT image into a one-sentence degradation observation plus an executable JSON command naming the degradation type, severity and tool to call. The controller is built on Qwen3-VL-4B and adapted with LoRA on attention and feed-forward projection layers at rank 16, scaling factor 32 and dropout 0.05, with the language-modeling loss computed only on the assistant response. On 2,000 patient-disjoint test images, the controller reached 95.7% accuracy for degradation type and 93.9% for severity, with a 100% JSON parse rate; most recognition errors fell between mild low-dose and clean images.

The sparse-view and metal-artifact specialists correct measurements in the projection domain, preserve acquired views through data consistency, then refine the image after differentiable filtered back-projection. A shared image-domain network cannot impose measurement consistency in the sinogram domain where the corruption actually occurs, whereas the dual-domain specialists do so explicitly. In the sparse-view ablation, the dual-domain specialist reached 46.04 dB with 4.2 million parameters, against 44.43 dB for single-task AMIR and 43.57 dB for all-in-one AMIR with 23.5 million parameters.

Experiments used the public, de-identified Mayo 2020 abdominal CT data with a patient-disjoint training-test split, and degradations were simulated, covering three low-dose levels, three sparse-view levels and titanium, iron and gold metal settings. Restoration quality was compared under routing with ground-truth degradation labels, while controller perception and routing were evaluated separately on 2,000 test images.

Overall restoration performance under automatic routing is not reported, and the main restoration comparison relies on ground-truth degradation labels. Each experiment contains one degradation at a time, and tool ordering and iterative restoration for scans with multiple simultaneous degradations are left unstudied. The method has also not been validated on independently collected real clinical degraded CT across scanners, institutions and acquisition protocols. The reported comparisons are descriptive, with no formal significance tests.