Tutorial
Tutorial: Machine Learning in Cardiovascular MRI II
ISMRM & SMRT Annual Meeting • 15-20 May 2021
Tutorial
Tutorial: Machine Learning in Cardiovascular MRI II
| Concurrent 7 | 14:00 - 15:00 | Moderators: Sebastian Kozerke & Matthan Caan |
You will find the tutorial spaces to your left as you enter the Mount Seymour lobby. The awards will be presented in the Discovery Lecture Hall.
At the beginning of the session, please go to the Discovery Lecture Hall and join the Zoom call for instructions. You can join the zoom call by clicking "x" as soon as you enter the space. Please note: in some browsers it may be necessary to mute yourself in gather.town as you join the zoom call.
Gather.town will provide more options for small group discussion and interaction, but the presentations can also be viewed live outside of gather.town by following a Zoom link in the session page on Pathable.
Overview
In these tutorials, the audience will interactively participate in developing algorithms for machine learning of three tasks:
1. Automatic segmentation of the aorta on 4D flow MRI data;
2. Learning to map genetics onto the heart; and
3. Disentangled representation learning in cardiac image analysis.
Target Audience
Clinicians and engineers interested in machine learning.
Educational Objectives
As a result of attending this course, participants should be able to:
- Explain the basics of machine learning;
- Identify machine learning algorithms for different tasks; and
- Recognize and be able to design code for machine learning problems.
| Disentangled Representation Learning in Cardiac Image Analysis
Thomas Joyce
Machine learning has long been concerned with the idea of “representation learning” - how can algorithms learn good ways to represent data, such that these representations make downstream tasks easier. This tutorial will be a ‘hands on’ introduction to learning low dimensional representations of cardiac MR images using python [1] and pytorch [2]. Specifically, we will explain and implement a variational auto encoder [3] (VAE, https://gitlab.ethz.ch/joycet/cardiacvae), and demonstrate how a VAE can be used to learn low dimensional (vector) representations of MR images. We will also explore using the same VAE to learn representations of the anatomical masks corresponding to the data. Once the VAE is in place we will then examine what has been captured in the dimensions of the learned latent space. Finally, we will also demonstrate the potential for using the VAE models to generate new synthetic MR and anatomy images. Though-out, we will use images from open cardiac MRI datasets (from ACDC [4] and Kaggle [5]), and the code will also be available, so that participants can try the experiments out themselves, and explore their own modifications and improvements. |
|||
| Learning to Map Genetics onto the Heart in Large Populations
Declan O'Regan
In this talk, participants will be given a background to clinical cardiac MRI and the motivation for developing machine learning approaches for segmentation, classification and prediction using complex traits. This will focus on atlas-based approaches for image registration, 3D regression modelling, and using autoencoders for disease classification and prediction tasks. Lastly, we will focus on two examples of how imaging can be integrated with genetic data. The first will be modelling how rare genetic variants, associated with heart failure, affect the structure of the heart in apparently health adults but still predispose to disease. The second example will be using high-throughput image analysis in large genotyped populations to understand the genetic regulation of structural complexity – as captured using MRI. This will give participants an understanding of the downstream assessments of imaging data that are possible using quantitative analysis of complex motion and geometrical traits in biomedical science. Further reading: |
|||
| Machine Learning Segmentation of Heart/Aorta
Jelmer Wolterink
In this tutorial, participants will be introduced to deep learning basics for the segmentation of cardiac cine MR images using publicly available software tools and a public challenge data set. Specifically, participants will learn how to train a U-Net neural network architecture [1] to segment the left ventricle myocardium and cavity, and right ventricle, in short-axis cardiac cine MR images using the Python [2] programming language and the popular PyTorch [3] and MONAI [4] frameworks for deep learning. The data set used is the training set of the Automatic Cardiac Diagnosis Challenge (ACDC) that participants can already download prior to the workshop from the challenge website [5,6]. This data set contains cardiac cine MR images and manual reference annotations of cardiac structures. The first session will provide a brief introduction to the clinical task and its relevance, the data set used, and programming components needed to tackle this problem using deep learning. Instructions will be provided for the data set and the programming and computing tools needed. Participants are encouraged to already install JupyterLab [7] on their own machine prior to the workshop. Alternatively, Google Colab [8] can be used for cloud computing instead of local computing. A scaffold of code will be provided for data loading and visualization, and model training. Participants are invited to extend this scaffold during the conference and train and evaluate their own deep learning model. The second session will reflect on one possible implementation of the deep learning approach and its results. Participants are kindly invited to share their approach or results before this meeting to be included in this presentation. |
The International Society for Magnetic Resonance in Medicine is accredited by the Accreditation Council for Continuing Medical Education to provide continuing medical education for physicians.
Back to Meeting Home
Back to the Program-at-a-Glance