Target Audience

Scientists and clinicians who are familiar with MR spectroscopy and would like to learn more about how deep learning algorithms can enable them to better understand the data they collect.

Learning Objectives

1. Understand how the structured nature of MR spectroscopy makes it amenable to certain deep learning techniques.
2. Know the difference between supervised and unsupervised deep learning techniques, and how to structure the input and output for each of these types of algorithm.
3. Identify three components of the MR spectroscopy processing pipeline in which deep learning can be of use and look at representative examples of literature for each: reconstruction, quantification/post-processing, and quality control.

Outline of Topics

First, the difference between supervised and unsupervised deep learning will be explored. Supervision refers to how a machine learning algorithm is trained, e.g. is it guided by known output (supervised), or can the output be inferred from some properties of the input (unsupervised).

Second, the inherent structure of MR spectroscopy data will be explored, and how deep learning algorithms can take advantage of this information in order to better identify relationships between data input and the desired output.

Finally, a literature review of deep learning of MR spectroscopy will be presented. There are three core areas in which papers fall, each covering one aspect of MR spectroscopy data processing. Representative papers from each will be covered in brief, with a list of additional resources presented at the end of each section. The core areas are:
1. Reconstruction – the goal of this first step is to convert acquired signals (e.g. free induction decays) into usable data. With the massive size of data in MR spectroscopy (e.g. 10-100,000 voxels in a single field-of-view), techniques such as undersampling and compressed sensing seek to reduce acquisition time and data size at the cost of needing to recover the missing data. Recent work into CNNs to perform reconstruction will be explored.
2. Quantification / Post-Processing – the process of quantifying MR spectroscopy data includes identifying the presence and concentration of metabolites and macromolecules within a voxel. In general, there have been two tasks that this body of work has tried to solve: determination of metabolite concentrations (relative concentrations), or actual fittings that include a model of the baseline and each metabolite. Extensive work over the past two decades has been performed, and a sample of the different deep learning architectures (both supervised and unsupervised) will be discussed.
3. Quality Control – there are many factors that can affect the usability of MRS data. While the gold standard is manual review by an expert, this does not scale well with the ever-increasing size of data. With deep learning, a new set of QA algorithms developed over the past decade seek to generate tools that can mimic the expert’s gestalts.

Acknowledgements

No acknowledgement found.