Case Western Proves Its DeepONet Works Before Showing It Works in Medical Imaging
Researchers prove that Deep Operator Networks can invert the Neumann-to-Dirichlet map in Electrical Impedance Tomography, with universal approximation guarantees for operator-to-function learning.
Takeaways
- Paper extends DeepONets to operator-to-function maps, targeting the Electrical Impedance Tomography inverse problem.
- Establishes a universal approximation theorem for approximating the map from Neumann-to-Dirichlet operators to conductivity functions.
- Splits error into encoder, approximator, and reconstructor terms, each bounded independently using Calderon problem stability.
- Architecture uses ResNet-based branch and trunk networks combined via inner product to output pointwise conductivity.
- Numerical study compares against Iteratively Regularized Gauss-Newton on 16,384 phantoms with 1 to 9 percent noise.
- Code available from authors on request; framework generalizes to other operator-input inverse problems.
A new paper from Anuj Abhishek (Case Western Reserve University) and Thilo Strauss (Xi'an Jiaotong-Liverpool University) does something unusual for deep learning applied to medical imaging: it proves the network works before showing that it works. The authors extend Deep Operator Networks (DeepONets) to the operator-to-function regime and establish a universal approximation theorem for reconstructing internal conductivity from boundary measurements in Electrical Impedance Tomography (EIT). The theoretical result is the headline, and the empirical benchmark against the classical Iteratively Regularized Gauss-Newton (IRGN) method is a supporting act.
The imaging problem hiding inside an operator
Electrical Impedance Tomography injects small currents through electrodes attached to a patient's skin, measures the resulting voltages, and reconstructs the internal conductivity map of the tissue. Clinical uses include lung function monitoring, breast cancer detection, cerebral edema detection, epilepsy seizure onset zone imaging, and bedside stroke monitoring. Compared to CT or MRI, EIT offers lower spatial resolution but faster acquisition, cheaper hardware, and no ionizing radiation.
Recovering conductivity from boundary measurements is highly nonlinear and severely ill-posed. Small measurement errors can blow up into large reconstruction errors, and regularization is needed to get anything sensible out. The mathematical object connecting the measurements to the physics is the Neumann-to-Dirichlet (N-t-D) operator, which maps applied boundary currents to measured boundary voltages. The inverse problem: given this operator, find the conductivity function inside.