Anatomy-guided AI reconstructs chest CT scans from sparse X-ray views

The method, developed by Ahmad Veysi, Farshid Babapour Mofrad, Hadi Zayyani, Hamidreza Pouraliakbar and colleagues, combines a U-Net neural network with a physics-based regularization scheme that adapts itself to the anatomy visible in each scan. Sparse-view CT reconstruction is a difficult inverse problem: when the number of projection angles is reduced to lower radiation dose or scanning time, images suffer from severe undersampling artifacts and amplified noise, and fine anatomical structures can be lost. Conventional regularization-based approaches rely on spatially uniform priors and often require manual tuning of parameters, while many deep learning methods offer limited interpretability and weak integration of physical constraints.

The new framework addresses both shortcomings by making the regularization itself spatially variant and guided by what the network recognizes in the image. At the core of the method is spatially variant Relative Total Variation (SvRTV) regularization, a mathematical prior that suppresses noise and texture-like artifacts while preserving true structural edges. Rather than applying this prior uniformly, the network predicts reconstruction parameters directly from image quality characteristics, including the overall regularization strength and tissue-specific weighting parameters designated A1 and A2.

In parallel, the same U-Net performs anatomical segmentation of the chest, and confidence maps derived from segmentation accuracy and edge information determine how strongly each tissue region is regularized. Training the network to perform segmentation, parameter estimation and reconstruction simultaneously raises the problem of balancing multiple competing objectives. The researchers solved this with a homoscedastic uncertainty-based multi-task learning strategy, which automatically weighs the different loss terms during training instead of relying on hand-tuned weights.

The result is a single model that adapts to varying acquisition conditions without needing to be retrained separately for each sparse-view configuration. In head-to-head comparisons, SvRTV-UNet was evaluated against conventional iterative reconstruction methods, model-based regularization approaches and representative deep learning baselines. The framework demonstrated improved reconstruction performance across quantitative metrics, and the authors report that qualitative results likewise supported its effectiveness relative to the evaluated approaches.

Because the predicted parameters are explicit rather than buried in the network's weights, the method offers a more interpretable reconstruction pipeline than purely black-box deep learning models. The study is limited to chest CT reconstruction, and the reported gains are relative to the specific methods and datasets used in the evaluation; the published version is an early-release article that remains subject to further editorial edits.

Subject of Research: Physics-guided deep learning for sparse-view chest CT reconstruction

Article Title: Anatomy-guided U-Net deep learning for sparse-view chest CT reconstruction using spatially varying RTV regularization

Article References: Veysi, A., Babapour Mofrad, F., Zayyani, H., & Pouraliakbar, H. (2026). Anatomy-guided U-Net deep learning for sparse-view chest CT reconstruction using spatially varying RTV regularization. Scientific Reports, 16(1), Article 34519. https://doi.org/10.1038/s41598-026-73560-1

Image Credits: AI Generated

DOI: 10.1038/s41598-026-73560-1

Keywords: computed tomography, sparse-view reconstruction, deep learning, U-Net, regularization, medical imaging, image segmentation, multi-task learning, chest imaging, radiation dose reduction, Anatomy-guided, deep