A transformer model designed to fill in missing projection views from sparse computed tomography scans posted the highest listed average scores in a fixed test and outperformed a comparison model when the number of views changed. In the fixed 64-to-256 setup, TomoTransformer scored 39.66 decibels on PSNR, one of the study's image-quality measures, and 0.931 on SSIM, another. Restormer scored 39.47 and 0.927. These figures were calculated against dense-view filtered back-projection references in computational experiments, so they describe reconstruction quality rather than a clinical outcome.
A model built for changing views
The central question was whether one transformer could work across different projection counts, angular arrangements, detector dimensions and image collections without being retrained. The proposed system represents each single-view filtered back-projection as an angle-conditioned local token, then predicts missing views in a shared, spatially aligned back-projection space. In the paper's framing, this representation separates scan geometry from image content.
Training drew on approximately 650,000 two-dimensional slices: about 570,000 CT slices from 1,340 volumes and 80,000 natural images. The evaluations simulated two-dimensional parallel-beam CT, reconstructed both sparse and dense inputs with filtered back-projection, and compared the outputs with dense-view targets using PSNR and SSIM. For the fixed benchmark, the competing methods were trained on the same heterogeneous dataset, while the model used 64 projections sampled uniformly from a 256-projection reference.
Testing beyond one protocol
The broader test changed the number and placement of views instead of holding one protocol constant. It used dense grids from 128 to 512 views, sparsity ratios from 20% to 80%, angular perturbations and native 362-by-362-pixel LoDoPaB-CT images. Against a dense 256-view reference, TomoTransformer scored 42.40 dB and 0.959 SSIM when 128 views were used to produce 256, compared with 37.75 dB and 0.914 for ViewTrans. With 64 input views and the same 256-view target, the scores were 37.68 and 0.910 for TomoTransformer, versus 33.07 and 0.832 for ViewTrans. The model therefore beat that comparator in both tested variable-view settings.
Noise and a real sample
The study also tested whether the model could handle noise it had not seen during training. A model trained on noiseless projections was applied to noisy sparse-view projections without noise training or fine-tuning. In the displayed condition, the input FBP reconstruction had a PSNR of 30.03 dB. The listed variants reached 35.49 dB using only predicted views, 35.09 dB using predicted plus measured views, 36.58 dB after blind-spot re-masking and 37.19 dB after complementary re-masking. The last of these was the highest displayed denoising score, but the result comes from selected simulated-noise conditions.
To test transfer beyond the simulations, the authors used ptychographic X-ray tomography of a mouse-brain pillar. The experiment had 617 views across 0 to 180 degrees, 37.6-nanometre pixels and a 768-pixel detector. The sparse input used 128 views, while a dense 512-view FBP served as the reference. Against that reference, TomoTransformer scored 35.8 dB and 0.837 SSIM when it generated a 256-view reconstruction, and 38.8 dB and 0.916 when it generated 512 views. Sparse FBP scored 31.8 dB and 0.722. On this sample, asking for more target views was associated with higher reported scores.
What the evidence does not establish
The authors report that TomoTransformer also outperformed the protocol-specific baselines on the real projections. They say those baselines hallucinated texture and left severe blocking artifacts, although the supplied comparison does not include the full numerical results for that statement. The real test was one mouse-brain pillar, and its benchmark used a dense FBP image rather than independent ground truth. The result therefore does not establish whether the model recovered true structure or would improve a clinical scan.
The paper is an arXiv preprint, version 2 dated 2 October 2026. Its evidence is primarily computational, with simulated two-dimensional parallel-beam benchmarks and one real nanoscale mouse-brain pillar. PSNR and SSIM are image-quality surrogates, and the study does not establish diagnostic accuracy, patient outcomes, performance on cone-beam CT or broad three-dimensional clinical deployment. The reported comparisons are averages, and no confidence intervals, p-values or repeated-run variability are given. Further testing would need additional real acquisition geometries, more varied noise and artifacts, and evaluation beyond the two image-quality scores.
Paper data and sources
Original title: TomoTransformer: Towards a Foundation Model for CT Reconstruction
Authors: AmirEhsan Khorashadizadeh, Benjamín Béjar
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text