Research / Published
AbdomenGen
How can we vary organ size while keeping a generated abdomen anatomically coherent?
On this page
The problem
Virtual imaging studies need anatomical variation that can be controlled and interpreted. Changing an organ’s volume should not require giving up the spatial relationships that make an abdomen plausible.
The computational idea
AbdomenGen uses volume-conditioned diffusion to generate abdominal anatomy sequentially. Three-dimensional signed-distance fields represent organ shape; conditioning on the body and previously generated structures gives each step anatomical context.
The Volume Control Scalar separates organ size from body habitus, making the requested change easier to interpret than raw volume alone.
Research context
This is my first-author MICCAI 2026 work at Duke’s Center for Virtual Imaging Trials, with Lavsen Dahal, Paul Segars, and Joseph Y. Lo.
Evaluation and limitations
The paper evaluates anatomical fidelity, volume calibration, and multi-organ control. Controllable anatomy is useful for virtual imaging studies, but is not itself evidence of clinical validity. The full paper describes the evaluation conditions and limitations; this page does not substitute illustrative geometry for experiment figures.