New AI Framework Bakes Anatomy Rules Into Neural Networks

A new paper posted to arXiv proposes a way to stop deep-learning models of human anatomy from generating results that look statistically reasonable but are physically impossible. The framework, called Anatomy-Informed Neural Networks (AINN), was authored by David Stonko and described in a 42-page paper with 10 figures and 4 tables, as reported on arXiv (cs.AI, cross-listed to cs.CV and cs.RO).

The core problem the paper addresses is that deep-learning anatomy models "can be numerically plausible yet anatomically impossible," and they tend to generalize poorly when training data is limited — a common situation in medical imaging. AINN addresses this in two ways. Soft anatomic priors are added as penalty terms in the loss function, in direct analogy to physics-informed neural networks (PINNs), which already penalize violations of physical laws during training. For example, a renal transplant artery branching off the iliac artery instead of the aorta would be flagged as unusual rather than treated as impossible. Hard anatomic priors — such as the requirement that a blood vessel remain continuous — are instead built directly into the model's architecture and state representation, making certain invalid predictions structurally impossible wherever the constraint allows it.

To demonstrate the framework, the authors tackle a clinical scenario with very limited available data: predicting how the aortoiliac artery tree deforms when a stiff wire is threaded inside it during a procedure, a mechanical effect relevant to aortic surgery and to future autonomous endovascular robots that would need to navigate vessels on their own. The method represents the vessel's centerline and the wire's path not as simple 3D curves but as curves of reference frames in SE(3), the mathematical group used to describe combined position and orientation in three dimensions. A Cosserat-rod model — a physics formulation capable of representing a thin elastic rod's bending and twisting — is used for the wire, coupled to a vessel model that accounts for its natural tortuosity and anatomical anchoring. The two interact through a one-directional contact constraint representing the wire pressing against the inner vessel wall, and the predicted deformation is computed as the solution that minimizes the combined elastic energy of the system, with contact forces emerging as Lagrange multipliers of that constraint.

Training supervision comes from comparing a 2D projection of the predicted 3D deformation, generated using the geometry of a C-arm X-ray imaging system, against an actual angiogram image, using a Wasserstein-2 optimal-transport loss to measure the difference. This allows a single 2D angiogram to supervise a 3D prediction.

Importantly, the authors note that in this paper no neural network is actually trained; the current work verifies the kinematics, loss formulation, and projection geometry against known ground truth, while the mechanics solver has only been checked against its own internal optimality conditions, and the predicted displacement has not yet been shown to be mesh-converged. Future work is planned to apply the in silico model to real CT scan data to test whether it improves prediction accuracy and reduces the amount of training data required — a step that would need to be validated before any clinical or robotic navigation application.