Imaging Physics meets Machine Learning: AI approaches for Image Reconstruction and Acquisition

Over the past few decades, top-down expert engineering has driven the
creative design of tomographic imaging acquisition and reconstruction
processes. Image reconstruction is challenging because analytic
knowledge of the exact inverse transform may not exist a priori,
especially in the presence of sensor non-idealities and noise. Thus,
the standard reconstruction approach involves approximating the
inverse function with multiple ad hoc stages in a signal processing
chain whose composition depends on the details of each acquisition
strategy. We present here a unified framework for image
reconstruction, AUtomated TransfOrm by Manifold APproximation
(AUTOMAP), which recasts image reconstruction as a data-driven,
supervised learning task that allows a mapping between sensor and
image domain to emerge from an appropriate corpus of training
data. Implemented with a deep neural network, AUTOMAP is remarkably
flexible in learning reconstruction transforms for a variety of
acquisition strategies, utilizing a single network architecture and
hyperparameters. We further demonstrate its efficiency in sparsely
representing transforms along low-dimensional manifolds, resulting in
superior immunity to noise and a reduction in reconstruction artifacts
compared with conventional handcrafted reconstruction methods. In
this talk we also describe work deploying machine learning for MR
image acquisition with AUTOmated pulse SEQuence generation (AUTOSEQ),
using both model-based and model-free reinforcement learning
approaches to produce canonical (gradient echo) as well as
non-intuitive pulse sequences that can perform spatial encoding and
slice selection in unknown inhomogeneous B0 fields.