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.
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.
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.
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.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!