Tomo-center: AI finds rotation axis with sub-pixel accuracy for synchrotron tomography

Accurate rotation-axis alignment is critical for artifact-free reconstruction in parallel-beam synchrotron micro-tomography. Traditional methods like Vo's algorithm rely on sinogram features that fail on low-contrast or weakly absorbing samples. The new tomo-center method, presented by Songyuan Tang and colleagues, reframes center selection as a binary classification task. It uses a DINOv2-pretrained vision transformer with an attention-based multiple-instance learning head, fine-tuned end-to-end on tomographic images. At inference, the algorithm sweeps through candidate centers, reconstructs a stack of tomograms, and picks the optimal one.

The team tested tomo-center on two independent datasets, consistently achieving a mean absolute error below 1 pixel. Crucially, it remained robust under sparse or noisy acquisitions: performance held steady when the number of projections was reduced by a factor of 10, and when Poisson noise (via blank scan factor) was amplified 10-fold. Interpretability maps showed which continuous spatial features drive classification. The tool is now delivered as an open-source command-line tool, and has been integrated into several tomography software packages for routine beamline operations, promising sharper 3D reconstructions across biology, materials science, and geophysics.