Transformer AI Reconstructs Sharp CT Scans From a Fraction of the Radiation Dose

Computed tomography has become one of medicine’s most indispensable diagnostic tools, but every scan comes with a hidden cost: ionizing radiation. A new study published in Frontiers in Nuclear Medicine reports an artificial intelligence framework called Cross-Domain TransNet that can reconstruct diagnostically useful CT images from a small fraction of the usual number of X-ray projection views, potentially allowing clinicians to slash radiation exposure without sacrificing image quality. The work, led by a team of radiologists and engineers from hospitals and universities across China, tackles one of the most persistent problems in modern medical imaging: how to keep pictures sharp when the scanner collects far less data than normal.

The health stakes are considerable. The research team cites long-term epidemiological evidence linking CT radiation exposure to elevated cancer risk in both adults and children. One study of 31,462 patients found that CT-related exposure increased lifetime cancer risk by 0.7 percent across the population and overall cancer mortality by 1 percent, with risks climbing to between 2.7 and 12 percent in patients who underwent multiple scans. Children are particularly vulnerable, being three to four times more sensitive to radiation than adults, which can translate into more severe long-term health effects. These figures explain why the field has pushed hard toward sparse-view CT, an acquisition strategy that deliberately reduces the number of angular projections the scanner records, thereby lowering the X-ray dose delivered to the patient.

The catch is physics. When a CT scanner collects fewer projection views, the mathematical inversion problem that converts raw projection data, known as the sinogram, into an image becomes severely underdetermined. Applying the standard clinical workhorse, filtered back projection, to such sparse data produces images riddled with streak artifacts and noise that can render them clinically useless. Traditional iterative reconstruction methods that incorporate image priors such as total variation or dictionary learning can help, but they depend on iterative forward and backward projection operations, require detailed knowledge of the specific scanner geometry, and carry high computational complexity and long reconstruction times.

Deep learning has transformed this landscape over the past several years. Convolutional neural network approaches such as RED-CNN, which combines deconvolution, autoencoders, and residual connections, demonstrated that networks trained on large datasets could suppress noise and recover fine details that classical methods could not. But most of these image-domain methods share a fundamental limitation: they operate on images that have already been reconstructed from incomplete data, meaning information lost in the projection domain can never be recovered. As the authors note, this sequential strategy cannot recover information lost in projection data, limiting the potential of image-only approaches.

A second family of methods works in the projection domain, using neural networks to fill in missing sinogram data before reconstruction. Generative adversarial networks such as Pix2Pix have been used for pixel-level sinogram completion, and unsupervised projection-domain methods have achieved performance comparable to supervised approaches without requiring full-dose ground truth. These projection-domain techniques are more robust to noise, but they typically require detailed knowledge of the scanner’s hardware parameters, which vary widely across manufacturers and models, limiting practical applicability. Moreover, when deep learning is used only for projection repair, the limitations of traditional reconstruction algorithms still affect the final image.

Dual-domain methods attempt to get the best of both worlds by processing projection data and image data simultaneously. Representative systems include DuDoNet, which employs a mask pyramid U-Net for projection preprocessing and image post-processing, and DuDoTrans, which integrates Swin-Transformer and CNN components to repair sinograms. Yet the new study argues that most existing dual-domain frameworks still follow a sequential workflow: repair the projection data first, reconstruct an image with filtered back projection, then refine the image. This essentially couples two independent processes without fully exploiting the correlations between domains, and repeated use of filtered back projection may introduce additional noise and artifacts.

Cross-Domain TransNet, abbreviated CDTransNet by its creators, takes a fundamentally different approach by making the two domains talk to each other continuously throughout the network. Built on the Vision Transformer architecture, the model consists of three main parts: an initial patch embedding operation, a stack of Cross-Domain Transformer modules, and a patch merging operation. Inside each Cross-Domain Transformer block, tokens are split into image-domain and projection-domain representations, domain-specific self-attention captures long-range dependencies within each domain, and an explicit cross-attention mechanism exchanges information between them. This bidirectional design forces image-domain features to guide sinogram restoration and vice versa, a progressive multi-stage interaction the authors say is absent from earlier dual-domain transformers that apply self-attention to each domain separately and rely on implicit mapping.

The architecture also replaces the standard multilayer perceptron found in conventional Transformer blocks with a novel Convolution Fusion Layer, which enhances feature interactions and restores spatial continuity after attention operations. This matters because the fixed-size patch embedding used in Transformer encoders can produce mild patch-edge artifacts when anatomical structures span patch boundaries, particularly at higher sparsity levels. Training is guided by a dual-domain loss function comprising three components: image reconstruction loss, image perceptual loss, and reconstruction projection consistency loss, ensuring that the network preserves image details while remaining faithful to the measured projection data.

The experimental results are striking. Using the publicly available 2016 NIH-AAPM-Mayo Low Dose CT Grand Challenge dataset, with 2,378 three-millimeter slices from 10 patients at 256 by 256 resolution, the team simulated a fan-beam geometry with 800 detector elements and evaluated performance at 32, 64, 96, and 128 projection views. At the sparsest 32-view setting, CDTransNet achieved a peak signal-to-noise ratio of 43.03 and a structural similarity index of 0.980, compared with 12.18 and 0.348 for filtered back projection, 28.83 and 0.771 for RED-CNN, 33.62 and 0.886 for DuDoTrans, and 40.06 and 0.969 for CNCL. At 96 views, CDTransNet reached a PSNR of 44.85 and an SSIM of 0.985. Histogram-based analyses showed the method produced the narrowest error distribution centered around zero, indicating less reconstruction bias and more accurate pixel-wise recovery than competing approaches. Ablation experiments confirmed that removing the cross-attention module severely degraded performance, demonstrating that cross-attention is the key operation enabling meaningful dual-domain interaction.

The method also showed robustness across different reconstruction kernels and slice thicknesses, maintaining strong metrics on 1mmB30, 1mmD45, and 3mmD45 protocols in addition to the primary 3mmB30 dataset. The authors are candid about limitations: the model relies on high-quality paired training data that may be difficult to obtain in practice, the experiments used simulated projections because raw clinical projection data were not publicly available, and the dual-domain attention design demands roughly 12 gigabytes of GPU memory and about 356 seconds of training time per epoch, with around 40 million parameters, making it heavier than competitors. Future work could explore lightweight variants through knowledge distillation or model pruning. Still, the study represents a compelling proof of concept that Transformer-based cross-domain attention can extract diagnostically reliable images from dramatically under-sampled scans, pointing toward a future where CT imaging delivers the same answers at a fraction of the radiation burden.

Subject of Research: Transformer-based dual-domain deep learning for sparse-view CT image reconstruction

Article Title: Cross-Domain TransNet for sparse-view CT reconstruction

Article References: Wang, J., Zou, C., Yang, H., Liao, P., Mo, X., Gao, Y., Ning, Y., & Lai, J. (2026). Cross-Domain TransNet for sparse-view CT reconstruction. Frontiers in Nuclear Medicine, 6, Article 1711533. https://doi.org/10.3389/fnume.2026.1711533

Image Credits: AI Generated

DOI: 10.3389/fnume.2026.1711533

Keywords: CT reconstruction, sparse-view CT, deep learning, Transformer, dual-domain, cross-attention, radiation dose reduction, medical imaging, sinogram, NIH-AAPM Mayo dataset, artifact suppression, computational imaging