TomoTransformer turns sparse-view CT reconstruction into a single retraining-free foundation model via back-projection space
Synopsis
The authors introduce TomoTransformer, a transformer architecture that treats each local filtered projection as an individual token and predicts missing views via self-attention in a back-projection space, yielding a single foundation model that can process any number of input projections at arbitrary angular locations and detector dimensions and query any number of target angles without retraining; trained on a large-scale dataset spanning diverse medical CT anatomies and natural images, it significantly outperforms concurrent multi-purpose models like ViewTrans and matches or exceeds strong protocol-specific baselines on several sparse-view benchmarks, while showing robust zero-shot generalization on real experimental nanoscale brain data collected from an X-ray synchrotron.
Interpretation
Introduces a transformer architecture that treats each local filtered projection as an individual token and predicts missing views via self-attention in a back-projection space, making view interpolation geometrically well-posed and invariant to detector size. Relative to conventional supervised models that map FBP images or sinograms to clean reconstructions, this reformulates reconstruction as view interpolation, removing the binding to fixed projection counts, angles, and detector resolutions. The abstract provides an architectural description and a geometric argument ("separates projections across spatial locations, making view interpolation geometrically well-posed and invariant to detector size"), which is a methodological argument rather than ablation data.
A single foundation model can process any number of input projections at arbitrary angular locations and detector dimensions and query any number of target angles without retraining. Conventional models require retraining whenever projection counts and angles, detector resolutions, or data distributions change; this work replaces that protocol dependence with full agnosticism to the number of input and target projections. The abstract states this capability directly with "without retraining" and "fully agnostic to the number of input and target projections," without giving numerical projection counts or angular ranges.
Trained on a large-scale dataset spanning diverse medical CT anatomies and natural images, the model generalizes across anatomies, materials, and resolutions, and on several sparse-view benchmarks significantly outperforms concurrent multi-purpose models like ViewTrans while matching or exceeding strong protocol-specific baselines. Relative to concurrent multi-purpose models and protocol-specific baselines, this work pursues cross-protocol generality together with competitive reconstruction quality rather than tuning to a single sparse-view protocol. The abstract reports "Extensive evaluations on several benchmark sparse-view datasets" and gives comparison conclusions against ViewTrans and protocol-specific baselines, but does not list dataset names, metric values, or sample sizes in the abstract.
Demonstrates robust zero-shot generalization on real experimental nanoscale brain data collected from an X-ray synchrotron. Relative to reconstruction models evaluated only on benchmark datasets, this pushes validation to real experimentally acquired nanoscale data, pointing toward practical deployment. The abstract states "robust zero-shot generalization on real experimental nanoscale brain data collected from an X-ray synchrotron," a conclusion on real data; the abstract does not provide specific error metrics.
Perspective
The work targets sparse-view tomographic reconstruction, suited to deployment settings where projection counts, angular locations, and detector dimensions vary and where per-protocol retraining is to be avoided; its stated scope includes diverse medical CT anatomies, materials, and resolutions, as well as nanoscale brain data collected from an X-ray synchrotron. For a reader, this means shifting the reconstruction pipeline from "one model per protocol" toward "one model across protocols," and trying zero-shot inference directly on real experimental data. The abstract does not give boundary values for projection counts, angular ranges, detector dimensions, or the number of target angles, so the practically usable range must be read from the paper's experimental setup.
The abstract does not list specific benchmark dataset names, metric values, training data scale, or ablation results, so the magnitude of "significantly outperforms ViewTrans" and "matches or exceeds strong protocol-specific baselines" cannot be judged from the abstract; the error level and evaluation protocol of the zero-shot nanoscale brain result are also not described. In addition, the abstract does not address the computational cost of constructing the back-projection space, sensitivity to projection noise and geometric calibration error, or behavior at very low projection counts, which are open questions a reader should confirm in the paper before adopting the method.
