One frontal X-ray, eight lumbar and pelvic bones in millimeters: Raysemble hits 65.24% Dice without rigid alignment

A single frontal X-ray can now yield millimeter-scale 3D shapes and relative positions for eight bones (L1–L5, the sacrum and both hip bones) without first synthesizing a whole CT volume and segmenting it. Earlier single-view routes either reconstructed CT intensities and then applied a segmenter for bone labels, or relied on statistical shape models and dual-surface depth fitting to complete bone surfaces. On a test set of 43 CT scans from 23 held-out patients, each rendered into one simulated frontal radiograph, Raysemble averages 65.24% Dice, 9.09 mm HD95 and 3.32 mm ASSD before rigid alignment, with the highest Dice and the lowest surface distances among 15 configurations.

The method estimates each bone's shape and its position in physical space separately, then links them through camera geometry. A normalized shape can occur at different positions and physical sizes, and the relation between depth and magnification is what connects center to scale. Each 3D query projects into the single frontal image and samples both regional and full-frame features, and a shared decoder outputs normalized occupancy; predicted depth sets the bone center, the magnification change with depth sets the scale, and the result is resampled onto a fixed grid at roughly 1 mm spacing.

Under identical inputs, predicting bone occupancy directly outperforms synthesizing CT and then segmenting it. Both pipelines receive the same frontal image, the same FleXray probability maps, the same bone regions and the same camera parameters, and VolumeNeRF keeps its original generator and deformable attention. Averaged per patient, Raysemble's Dice is higher for 18 of 23 patients, by 5.86 points on average (95% CI 3.35–8.94); HD95 and ASSD are lower for all 23 patients, by 10.55 mm and 2.44 mm on average.

The demonstrated scope is 43 CT scans from 23 held-out patients, each rendered into one simulated frontal radiograph, plus two real pelvic frontal radiographs from a single hospital; the latter are exploratory, the 16-day case was used during development, and the examinations were not simultaneous.

The real-radiograph comparison covers only two cases, with X-ray–CT intervals of 16 and 70 days; clinical evaluation before alignment requires calibrated camera geometry and confirmed image pairs, magnification errors can be mistaken for bone-size errors, and truncated bones can displace the predicted region and the initial query grid.