Back to AI Research

AI Research

Anatomy-Informed Neural Networks: Encoding Anatomic... | AI Research

Key Takeaways

  • Anatomy-Informed Neural Networks (AINN) are a proposed framework designed to improve the reliability of medical AI by embedding anatomical knowledge directly...
  • Deep-learning models of anatomy can be numerically plausible yet anatomically impossible, and they generalize poorly when data are scarce.
  • We develop it on a clinical test case with limited data: how the aortoiliac tree deforms when a stiff wire is introduced endoluminally.
  • This is important to contemporary aortic surgery and will matter to autonomous endovascular navigation.
  • The prediction is a constrained minimizer of the coupled elastic energy, with contact forces as its Lagrange multipliers.
Paper AbstractExpand

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 endoluminally. This is important to contemporary aortic surgery and will matter to autonomous endovascular navigation. We lift the vessel centerline and the wire path from R^3 to curves of frames in the Lie group SE(3), and couple a Cosserat-rod wire to a tortuosity-modulated, anatomically anchored vessel through a unilateral lumen-contact inequality. The prediction is a constrained minimizer of the coupled elastic energy, with contact forces as its Lagrange multipliers. Supervision is a Wasserstein-2 optimal-transport loss between the predicted projection through the C-arm geometry and the observed angiogram, so a 2D angiogram can train a 3D prediction. The kinematics, loss and projection are verified against known ground truth; the mechanics solver only against its own optimality conditions, and predicted displacement is not yet mesh-converged. Here, no network is trained. 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.

Anatomy-Informed Neural Networks (AINN) are a proposed framework designed to improve the reliability of medical AI by embedding anatomical knowledge directly into a model’s structure and training process. Developed by David P. Stonko, this approach addresses the tendency of deep-learning models to produce anatomically impossible predictions—such as vessels branching from the wrong location—when working with limited clinical data. By treating anatomical rules as either hard constraints or soft penalties, AINN aims to ensure that model outputs remain physically and biologically plausible. Methods and results are detailed in the full paper on arxiv.org.

The AINN Framework

The AINN paradigm functions similarly to Physics-Informed Neural Networks (PINNs). While PINNs use governing physical equations as penalty terms in a loss function, AINN uses anatomical priors. These priors are categorized into two types:

  • Hard Priors: Built directly into the model’s architecture and state representation. These make invalid predictions impossible by construction. For example, representing vessel centerlines as frames in the Lie group SE(3) ensures that every predicted configuration is a valid, rigid-body movement.

  • Soft Priors: Included as penalty terms in the loss function. These discourage anatomically unlikely configurations without strictly forbidding them, allowing the model to accommodate natural anatomical variations, such as rare branching patterns in patients with transplants.

Clinical Application: Vascular Deformation

The paper demonstrates this framework by modeling how the aortoiliac tree deforms when a stiff guidewire is inserted during endovascular surgery. This is a significant challenge because the preoperative CT scan (State 1) does not reflect the deformed state (State 2) caused by surgical instruments. To solve this, the model uses:

  • SE(3) Representation: The vessel centerline and wire path are lifted to the Lie group SE(3) to maintain geometric validity.

  • Cosserat-rod Coupling: A mechanical model couples the stiff wire to the vessel, accounting for lumen contact and the vessel's resistance to deformation.

  • Optimal-Transport Loss: The model uses a Wasserstein-2 loss to compare the predicted 3D deformation against a single 2D intraoperative angiogram, allowing 2D data to supervise 3D predictions. The Robotics story also surfaces in NVIDIA Launches Cosmos 3 Edge for..., adding another angle. The same Robotics question is explored in MNIST-PRO, which adds a research perspective.

Verification and Limitations

The study provides a numerical verification of these components using synthetic problems with known ground truth. The author notes that the current mechanics solver is verified against its own optimality conditions, but the predicted displacement is not yet mesh-converged. Furthermore, no neural network was trained in this specific study; the work establishes the framework and the mathematical foundation for future implementation. The architecture story also surfaces in New AI Architecture Mimics the Human..., adding another angle.

Why This Matters Current surgical planning often relies on static preoperative images, which can lead to geometric errors when the anatomy shifts during a procedure. By creating a model that respects anatomical constraints, AINN could improve the accuracy of surgical planning and support the development of autonomous endovascular navigation. The framework is designed to be data-efficient, which is critical in medical settings where large, annotated datasets are often unavailable. Future work will focus on applying this model to real CT scans to test its predictive accuracy in clinical environments.

More from Franklin

MNIST-PROresearchSymFoldresearchTrace Integrity for LLM Data AgentsresearchNew AI Architecture Mimics the Human...newsNVIDIA Launches Cosmos 3 Edge for...newsBlack Forest Labs Unveils FLUX 3...news

Comments