Curating large labeled CT datasets is expensive and limited by patient privacy. To address this, researchers introduce NV-Generate-BodyMask, a 3D latent diffusion model that synthesizes whole-body CT segmentation masks. The model allows precise, joint control over anatomical structures, tumor presence, and demographic characteristics, according to the abstract.

The approach could reduce the need for real patient data in training medical imaging AI, as synthetic masks can be generated on demand with desired properties. This may help overcome both cost and privacy barriers in dataset curation.

The work is described in an arXiv preprint and has not yet undergone peer review. Further details on the model architecture and evaluation are not included in the abstract.