Anatomy-Informed Neural Networks: Encoding Anatomic Priors in Loss and Architecture, with an SE(3) Formulation of Guidewire-Induced Aortoiliac Deformation
The paper proposes anatomy-informed constraints, but reports that no neural network has been trained yet.
David P. Stonko frames AINN as a way to put anatomic rules into both model loss and architecture, separating soft penalties from hard constraints. The test case is guidewire-driven deformation of the aortoiliac tree, modeled with SE(3) vessel and wire paths, elastic energy, and lumen-contact constraints. A 2D angiogram is used as supervision for a 3D prediction through a Wasserstein-2 projection loss. The paper verifies kinematics, loss, and projection against ground truth, while noting the mechanics solver is only checked against optimality conditions and displacement is not mesh-converged. ArXiv · AI/CL/LG's note
David P. Stonko frames AINN as a way to put anatomic rules into both model loss and architecture, separating soft penalties from hard constraints. The test case is guidewire-driven deformation of the aortoiliac tree, modeled with SE(3) vessel and wire paths, elastic energy, and lumen-contact constraints. A 2D angiogram is used as supervision for a 3D prediction through a Wasserstein-2 projection loss. The paper verifies kinematics, loss, and projection against ground truth, while noting the mechanics solver is only checked against optimality conditions and displacement is not mesh-converged. ArXiv · AI/CL/LG's note
score 4