Preprint: Topogram-Gated 3D U-Net Hits 0.005656 mumap MAE on BIC-MAC PET/MR Leaderboard

The arXiv preprint describes an ensemble of two complementary residual 3D U-Nets trained jointly in CT and 511-keV attenuation-map space. One model prioritizes cranial activity; the second adds a projection-domain attenuation-correction-factor regulariser. Equal averaging in CT space produced the submitted result. The topogram encoder supplies bounded multi-scale gating that injects global radiographic anatomy into the volumetric network.

This approach directly targets the clinical bottleneck of MR-based attenuation correction in whole-body PET/MR, where Dixon sequences alone miss bone and lung interfaces. By adding a low-dose topogram, the method supplies the missing projection constraint without requiring additional CT. Performance on organ bias (2.65 %) and brain outliers (0.0294) suggests the gating reduces both systematic and focal errors that currently degrade quantitative PET.

The work remains a single-team preprint submitted to an ongoing challenge; no independent replication or prospective patient outcome data exist. External validation on scanner vendors outside the BIC-MAC cohort and prospective SUV reproducibility studies are required before routine deployment.

If the reported error profiles hold on the final test set and subsequent multi-center cohorts, the technique could eliminate the need for separate CTAC scans in PET/MR workflows within three years.