Synopsis

Keywords: Image acquisition: Machine learning

This educational talk will present core concepts in deep learning, including convolutional neural networks (CNNs), batch-normalization, and loss functions. Next, it will review architectures such as encoder-decoder, GANs, and physics-guided unrolled neural networks. Finally, it will present guidelines for careful design of training databases and discuss caveats of current deep learning techniques.

Overview

Deep learning (DL) techniques are continuously making a high impact on the MRI field, improving tasks ranging from image reconstruction to data annotation and classification. This talk will introduce core concepts employed in DL, including convolutional layers, batch normalization, dropout, activation functions, loss functions, and optimization techniques. Then, we will discuss architectures that are useful for image reconstruction, e.g., encoder-decoder, generative adversarial networks (GANs), variational networks, and physics-guided unrolled networks. We will also discuss how to carefully choose training data from online public databases, in a task-adequate manner. Finally, we will discuss limitations and caveats of current techniques.

Acknowledgements

This work was supported by the Weizmann Institute Women’s Postdoctoral Career Development Award in Science.