Anatomy-Informed Neural Networks: Encoding Anatomic Priors in Loss and Architecture (arXiv 2608.21332)

Forum topic · 小凯 · 2026-08-25

Summary

Anatomy-Informed Neural Networks (AINN), introduced by David P. Stonko (arXiv:2608.21332), address a key weakness of deep-learning models of anatomy: predictions that are numerically plausible yet anatomically impossible, and poor generalization under data scarcity. AINN encodes soft anatomic priors 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 invalid predictions are impossible by construction. The method is developed on a data-limited clinical problem: predicting how the aortoiliac arterial tree deforms when a stiff guidewire is introduced transluminally. Vessel centerlines and wire paths are lifted into Lie group SE(3) frame curves, coupling a Cosserat-rod wire model with anatomy-anchored vessels via unilateral contact inequalities; predictions are constrained minimizers of coupled elastic energy with contact forces as Lagrange multipliers. Supervision uses a Wasserstein-2 optimal transport loss between predicted projections and observed angiograms, enabling 3D training from 2D images. The framework targets aortic surgery and autonomous endovascular navigation.

Overview

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

Summary (translated from the Chinese forum post)

Deep-learning models of anatomy can be numerically plausible yet anatomically impossible, and they generalize poorly when data are scarce. The paper introduces Anatomy-Informed Neural Networks (AINN), where:

Soft anatomic priors enter the loss function as penalty terms (e.g., a branching penalty that treats a renal transplant artery originating off the iliac rather than the aorta as *unexpected* rather than *impossible*), in direct analogy to physics-informed neural networks.
Hard anatomic priors (e.g., continuity of a vessel) are built into the architecture and state representation, making such invalid predictions impossible *by construction* wherever the prior admits architectural enforcement.

The framework is developed on a clinical test case with limited data: how the aortoiliac tree deforms when a stiff guidewire is introduced transluminally. This is relevant to contemporary aortic surgery and will also impact autonomous endovascular navigation.

Technical approach

Vessel centerlines and guidewire paths are lifted from R³ into frame curves in the Lie group SE(3).
A Cosserat-rod guidewire model is coupled with anatomy-anchored vessels through unilateral lumen-contact inequalities, modulated by torsion.
Predictions are constrained minimizers of coupled elastic energy, with contact forces acting as Lagrange multipliers.
Supervision uses a Wasserstein-2 optimal transport loss between predicted projections (via C-arm geometry) and observed angiograms, so 2D angiography can train 3D predictions.

Validation status

Kinematics, loss, and projection are validated against known ground truth.
The mechanics solver is validated only against its own optimality conditions; predicted displacements are not yet mesh-converged.
No network has been trained yet.

Future work

Transferring the in silico model to real CT scans and testing whether it improves predictive accuracy while reducing required training data.