DeVAR uses visual autoregressive modeling to denoise low-dose CT scans

CT scans are essential for diagnosis but radiation exposure remains a concern. Low-dose CT (LDCT) reduces risk but introduces noise and artifacts. Existing deep learning methods help, but they struggle with preserving fine anatomical details. Researchers from Shanghai Jiao Tong University introduce DeVAR, a generative framework that applies visual autoregressive modeling (VAR) to LDCT denoising for the first time. DeVAR conditions on global context from LDCT prefix tokens and progressively generates discrete token maps of normal-dose CT via next-scale prediction. Because quantization can discard high-frequency information, DeVAR includes a residual refiner to capture subtle anatomical structures beyond the codebook's capacity.

A dual-representation hybrid training strategy seamlessly integrates continuous and discrete latents, enabling the model to reconstruct high-fidelity, detail-preserved images. The system was evaluated on two public datasets, consistently achieving superior qualitative and quantitative performance compared to current state-of-the-art LDCT denoising methods. This approach demonstrates that autoregressive modeling, originally popular in text and image generation, can be effectively adapted for medical imaging tasks that require global-to-local structural awareness. The work opens a new path for safer, lower-dose CT protocols without compromising diagnostic accuracy.