Machine Learning Tutorial Session
New for 2024 the conference program will conclude with the Tutorial Session on Thursday, April 11, 4:00 pm. Join us to learn about THE hot topic of machine learning and its use cases for NMR.
Two complementary lectures will provide an intro and background on machine learning (lecture 1) and then demonstrate use cases and applications for magnetic resonance (lecture 2).
Lecture 1: Juan Eugenio Iglesias
Associate Professor of Radiology
Machine Learning for Magnetic Resonance: From Bayesian Methods to Modern Neural Networks for Large-Scale Analysis of MRI Data
Every year, huge amounts of brain MRI scans are acquired at hospitals for clinical purposes. These scans could be used for neuroimaging studies with sample sizes much larger than those found in research, but one needs to tackle the extreme heterogeneity of these data in terms of image appearance (MRI pulse sequence, resolution, etc). In this talk, I will first introduce classical Bayesian segmentation techniques that were introduced decades ago. Then, I will summarize how these methods where superseded by modern machine learning techniques (particularly deep convolutional neural networks), which provided excellent performance but little robustness to variations in image appearance. Finally, I will present a family of machine learning techniques developed by my group that solve this problem, along with results on clinical datasets with >10,000 scans.
Lecture 2: D. Flemming Hansen
Professor and Chair of NMR Spectroscopy
Transforming and Analyzing Complex NMR Spectra with Deep Neural Networks
Artificial intelligence (AI) and deep learning are now established as some of the most important technologies of our time, however, the uptake of AI in nuclear magnetic resonance (NMR) spectroscopy has been slower - but this picture is now swiftly changing. In this talk, I will first show examples of how deep neural networks (DNNs) can be trained to transform biomolecular NMR spectra, including DNNs for virtual homonuclear decoupling, reconstruction of sparsely sampled spectra, and resolution enhancements. Another area that will be presented is developments for autonomous analysis of complex NMR data. As an example, DNNs for the analysis of 1H chemical exchange saturation transfer (CEST) data, will be discussed, where the DNN not only accurately predicts the chemical shifts of nuclei in the exchanging species, but it also determines the uncertainties associated with these predictions. Common to applications of DNNs in NMR is that they do not contain any parameters for the end-user to adjust and the methods therefore allows for autonomous analysis.