Anatomy-Informed Neural Networks: Encoding Anatomic Priors in Loss Functions and Architecture

Forum topic · 小凯 · 2026-08-25

Summary

This arXiv paper (2608.21332) by David P. Stonko introduces Anatomy-Informed Neural Networks (AINN), a framework that embeds anatomical knowledge into deep learning models. Soft anatomic priors act as penalty terms in the loss function—analogous to physics-informed neural networks—while hard priors, such as vessel continuity, are built into the architecture so that invalid predictions are impossible by construction. The work is developed on a data-scarce clinical case: predicting how the aortoiliac arterial tree deforms when a stiff guidewire is introduced transluminally, relevant to aortic surgery and autonomous endovascular navigation. Vessel centerlines and guidewire paths are lifted from R3 to frame curves in the Lie group SE(3), coupling a Cosserat-rod guidewire model with anatomy-anchored vessels via unilateral lumen-contact inequalities; predictions are constrained minimizers of coupled elastic energy with contact forces as Lagrange multipliers. Training uses a Wasserstein-2 optimal transport loss between C-arm-projected predictions and observed angiograms, enabling 3D prediction from 2D angiography. Kinematics, loss, and projection are validated against ground truth; the mechanics solver is validated only against its own optimality conditions. No network is trained yet; future work will transfer the in silico model to real CT scans.

Paper Overview

Field: Computer Vision Author: David P. Stonko Posted: 2026-08-21 arXiv: 2608.21332

Abstract (Translated)

Deep-learning models of anatomy can be numerically plausible yet anatomically impossible, and they generalize poorly when data are scarce. We introduce Anatomy-Informed Neural Networks (AINN), in which soft anatomic priors enter as penalty terms in the loss (e.g., a branching penalty that treats a renal transplant artery off the iliac instead of the aorta as unexpected rather than impossible), in direct analogy to a physics-informed neural network, and hard anatomic priors (e.g., continuity of the vessel) are built into the architecture and state representation, making such invalid predictions impossible by construction wherever the prior admits architectural enforcement. We develop it on a clinical test case with limited data: how the aortoiliac tree deforms when a stiff wire is introduced transluminally. This is important for contemporary aortic surgery and will also inform autonomous endovascular navigation.

We lift vascular centerlines and guidewire paths from R³ to frame curves in the Lie group SE(3), coupling a Cosserat-rod guidewire with anatomy-anchored vessels modulated by torsion through unilateral lumen-contact inequalities. Predictions are constrained minimizers of the coupled elastic energy, with contact forces as their Lagrange multipliers. Supervision is a Wasserstein-2 optimal transport loss between the predicted projection (via C-arm geometry) and the observed angiogram, so 2D angiograms can train 3D predictions. Kinematics, loss, and projection are validated against known ground truth; the mechanics solver is validated only against its own optimality conditions, and predicted displacements are not yet grid-converged. No network is trained here. Future work will transfer this in silico model to real CT scans and test whether it improves predictive accuracy and reduces the training data required.

Key Contributions

Soft priors: anatomic plausibility encoded as loss penalty terms, in direct analogy to physics-informed neural networks (PINNs)
Hard priors: structural constraints (e.g., vessel continuity) built into architecture and state representation, ruling out invalid predictions by construction
Clinical test case: aortoiliac tree deformation under stiff guidewire insertion, relevant to aortic surgery and autonomous endovascular navigation
Geometric formulation: centerlines and guidewire paths as frame curves in SE(3); Cosserat-rod mechanics coupled to vessels via unilateral contact inequalities with Lagrange-multiplier contact forces
2D-to-3D supervision: Wasserstein-2 optimal transport loss between C-arm-projected predictions and observed angiograms

Validation Status

Kinematics, loss, and projection validated against known ground truth
Mechanics solver validated only against its own optimality conditions
Predicted displacements not yet grid-converged; no network trained in this work
Future work: transfer the model to real CT scans and evaluate accuracy gains and data efficiency

--- *Auto-collected 2026-08-25*