A foundation model for three-dimensional anatomical reconstruction from sparse X-ray projections

Abstract Computed tomography (CT) provides three-dimensional anatomical information but requires costly scanners and exposes patients to higher radiation doses than conventional X-ray imaging. Reconstructing three-dimensional anatomy from a few X-ray projections could extend widely available X-ray systems beyond two-dimensional imaging. We present SPARC, a foundation model pretrained once on 47,149 chest scans across diverse acquisition settings, enabling a single model to be adapted to different X-ray imaging configurations and anatomical regions. Across eight datasets covering five body regions, seven mainly outside the chest, SPARC reconstructed volumes with greater fidelity than five existing methods, exceeding the strongest in peak signal-to-noise ratio by 0.8-2.6 dB with eight projections and 1.4-2.7 dB with four. CT-based tools for abnormality classification and organ segmentation retained most of their performance when applied directly to SPARC's reconstructions from eight projections, fewer than 3% of the hundreds recorded by a CT scanner. They reached 98% of their area under the receiver operating characteristic curve on CT for 18 chest abnormalities, and 86% and 93% of their Dice score on CT for abdominal organs and lung lobes, with Dice scores 8.8% and 4.4% higher, in relative terms, than with the best competing method. Pretraining considerably increased the adaptation data efficiency. With ten patients, SPARC reconstructed chest CT more accurately than the same architecture trained from scratch on 100 patients. With the models and code released, SPARC could support the clinical translation of sparse-view reconstruction and expand access to three-dimensional imaging where CT scanners are unavailable.