CardioComposer: Leveraging Differentiable Geometry for Compositional Control of Anatomical Diffusion Models
New framework uses 'differentiable geometry' to let researchers programmatically design realistic 3D organs for medical testing.
A research team from MIT, Harvard, and the Wyss Institute has developed CardioComposer, a novel AI framework that solves a critical problem in medical AI: the trade-off between geometric controllability and realism in 3D anatomical models. Traditional generative models for cardiovascular anatomy could create realistic structures but offered little control, or provided control at the expense of realism. CardioComposer bridges this gap by introducing a programmable, inference-time framework that generates multi-class anatomical label maps from simple, interpretable building blocks called ellipsoidal primitives.
These primitives represent fundamental geometric attributes—like the size, shape, and position of heart chambers or blood vessels. The key innovation is the use of 'differentiable geometry.' The team developed special measurement functions based on voxel-wise geometric moments that are mathematically differentiable. This allows the system to use loss-based gradient guidance during the sampling process of a diffusion model, meaning the AI can be steered with precise numerical constraints during image generation.
The result is unprecedented compositional control. Researchers can now disentangle and constrain individual geometric features—for instance, tweaking the thickness of a ventricle wall independently of its length—and combine multiple substructures programmatically. The method has been demonstrated to work across diverse anatomical systems, including cardiac, vascular, and skeletal organs, proving its versatility. The team has released their code publicly, paving the way for broader adoption in medical research and device development.
- Uses 'differentiable geometry' and gradient guidance to control diffusion models during image generation, enabling precise tuning of 3D structures.
- Generates models from interpretable ellipsoidal primitives, allowing programmatic control over size, shape, and position of anatomical substructures.
- Demonstrated effectiveness across cardiac, vascular, and skeletal systems, providing a versatile tool for clinical research and medical device testing.
Why It Matters
Enables rapid, cost-effective creation of accurate 3D organ models for testing medical devices and planning procedures, accelerating biomedical innovation.