Model-Based Supervised and Unsupervised Deep Learning

Direct inversion schemes that directly recover an image from undersampled measurements are widely used in imaging. By contrast, the model based image image reconstruction framework marries the physics of the acquisition with learnable deep learning modules. The formulation provides a systematic approach for deriving deep architectures for inverse problems with the arbitrary structure. This approach dramatically reduces the demand for training data and training time.

Acquiring fully sampled training data is often challenging in many applications, especially in ultra-high-resolution imaging and dynamic MRI. We introduce novel strategies including ensemble Stein's unbiased risk estimator (ENSURE) to train deep learning problems from undersampled data. Although our focus is on MRI, the approach is broadly applicable to other imaging problems (e.g. deblurring, inpainting), where uncorrupted data is not readily available.