Model Adaptation In Biomedical Image Reconstruction

Abstract

Deep neural networks have been successfully applied to a wide variety of inverse problems arising in biomedical imaging. These networks are often trained using a forward model, which is not only used to generate the training data but is often incorporated directly into the network itself. However, these approaches lack robustness to misspecification of the forward model: if at test time the forward model varies (even slightly) from the one the network was trained on, the network performance can degrade substantially. Given a network trained to solve an initial image reconstruction problem with a known forward model, we propose novel retraining procedures that adapts the network to reconstruct measurements from a perturbed forward model. Our procedures do not require access to more labeled data (i.e., ground truth images), but only a relatively small collection of measurements under the perturbed forward model. We demonstrate this simple retraining procedure empirically achieves robustness to changes in the forward model for image reconstruction in magnetic resonance imaging from undersampled k-space measurements.