MICCAI2021 Logo.png

Why we organize this tutorial:

Deep neural networks require a large amount of training data to generalize well. Unfortunately, large labeled datasets are difficult to obtain in the medical domain due to many reasons, including the privacy concern and sometimes the rare nature of the diseases.  This limitation has motivated the development of various transfer learning, domain adaptation, and domain generalization techniques to increase deep networks' data efficiency. However, these techniques assume the availability of data from all domains during the training phase. Moreover, they often require time-consuming offline model re-training, which leads to delays and disruption of clinical workflows.

Meta-learning, or learning to learn, has emerged as a promising paradigm to address data heterogeneity and scarcity in medical applications. It starts with the assumption that test data distribution is unseen and different from the training data. In contrast to learning a model from scratch (supervised learning) or transferring features (domain adaptation), meta-learning leverages knowledge from multiple tasks to improve the learning mechanism. Once learned, a meta-learning model can adjust itself to a new data distribution rapidly using only a few labeled examples. This technology's high computational and data efficiency potentially increases the accuracy, development speed, and deployment efficiency of computer-aided diagnosis systems.

While promising, meta-learning has been largely under-investigated in the MICCAI community. This tutorial aims to bridge the gap by providing a comprehensive introduction to meta-learning, including theories, applications, and future directions. The tutorial will invite leading researchers in meta-learning and medical imaging to present its state-of-the-art and explain how the techniques apply to medical image detection, classification, and segmentation tasks. Meta-learning workshops were held every year at Neural Information Processing Systems (NeurNIPS) conference from 2018 to 2020 and attracted large audiences. We expect our tutorial to bring this exciting and important research direction to MICCAI and attract a similar number of attendants. 


List of Objectives:

The tutorial is designed for machine learning researchers, biomedical engineers, medical practitioners, and graduate students who wish to learn the basics of Meta-Learning, state-of-the-art methods, and how to apply them to medical imaging applications. Specifically, the tutorial will help the audience achieve the following goals: 

  • Gain general knowledge about meta-learning technology

  • Understand basic mathematical formulation of meta-learning

  • Become familiar with the state of the art meta-learning techniques

  • Run popular meta-learning algorithms on various medical imaging applications

  • Know the current limitations and future directions of the field

Date and Time: TBD​

What will be covered:
The tutorial will last 8 hours, consisting of lectures, three keynote talks, invited talks on selected topics, and short coffee breaks. The structure is tentative as follows and will be subject to changes depending on real situations: 

  1. Opening remarks [5 minutes]

  2. Overview of meta-learning [30 minutes, Hien Nguyen] 

    • What is meta-learning?

    • Why is meta-learning useful for medical image analysis? 

    • Meta-learning taxonomy and comparison to domain adaptation, domain generalization, and transfer learning

  3. Meta-learning formal introduction [30 minutes, Hien Nguyen]   

    • Meta-learning terminology

    •    Meta-learning versus supervised learning

    •    Meta-learning from bi-level optimization and inference views

  4. Contemporary meta-learning algorithms [part 1, 60 minutes,  Swami Sankaranarayanan]

    • Metric learning: matching networks and prototypical networks [30 minutes]

    • Parameter initialization: model-agnostic meta-learning [30 minutes]

  5. Coffee break [10 minutes]

  6. Keynote speaker [60 minutes, Rama Chellappa]

    • Meta-learning for Computer Vision: Recent Developments and Remaining Challenges

  7. Keynote speaker [60 minutes, Nassir Navab]

    • Few-Shot Meta-Denoising of Medical Images

  8. Lunch Break

  9. Contemporary meta-learning algorithms [part 2, 60 minutes, Vishal Patel]

    • Learning to optimize  [30 minutes]

    • Memory-augmented meta-learning [30 minutes]

  10. Bayesian meta-learning [45 minutes, Swami Sankaranarayanan]

    • The variational lower bound (ELBO)

    • Amortized variational inference in Bayesian meta-learning     

    • Bayesian meta-learning evaluation

  11. Coffee Break [10 minutes]

  12. Keynote talk [60 minutes, Gustavo Carneiro]

    • Supervised and Unsupervised Task Design to Meta-Train Medical Image Classifiers

  13. Applications of meta-learning in medical domain [60 minutes]

    • Meta-learning for incidental lung nodule classification [20 minutes, Hien Nguyen] 

    • Active Bayesian meta-learning for brain cell classification [20 minutes, Vishal Patel] 

    • Meta-learning for disease identification from chest x-ray images [20 minutes, Ronald Summers] 

  14. Closing remarks [5 minutes]


About Organizers

  • Dr. Rama Chellappa Bloomberg Distinguished Professor in the Departments of Electrical and Computer Engineering and Biomedical Engineering at Johns Hopkins University (JHU). He is an expert in computer vision and machine learning with over 400 papers, 70,000 citations, and numerous awards. 

  • Dr. Ronald Summers is Senior Investigator, Chief of the Clinical Image Processing Service, National Institutes of Health. He has co-authored over 300 journal, review, and conference proceedings articles, and is a co-inventor on 12 patents. He is known for multi-organ multi-atlas registration and the development of large radiologic image databases including ChestX-Ray and DeepLesion databases.

  • Dr. Nassir Navab is a Professor and Chair of Computer Aided Medical Procedures & Augmented Reality, Department of Informatics at the Technical University of Munich. He is the author of hundreds of peer-reviewed scientific papers with more than 37000  citations. He is the author of more than thirty awarded papers including 10 at MICCAI and three at IEEE ISMAR.

  • Dr. Gustavo Carneiro is a Professor of the School of Computer Science at the University of Adelaide, ARC Future Fellow, and the Director of Medical Machine Learning at the Australian Institute of Machine Learning. He is an expert in computer vision, medical image analysis, and machine learning with over 150 published papers and 7,000 citations.

  • Dr. Hien Van Nguyen is an Assistant Professor of the Department of Electrical and Computer Engineering Department at the University of Houston. He has published 45 peer-reviewed papers and received 12 U.S. patents. He has successfully organized several MICCAI tutorials including deep learning for medical imaging (2015), deep reinforcement learning for medical imaging (2018), Bayesian deep learning (2019). 

  • Dr. Vishal M. Patel is an Assistant Professor in the Department of Electrical and Computer Engineering at Johns Hopkins University. He received the 2016 ONR Young Investigator Award, 7 paper awards, and 12000 citations. He is an Associate Editor of the IEEE Signal Processing Magazine, IEEE Biometrics Compendium.

  • Dr. Swami Sankaranarayanan is a Postdoc of the Department of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology

  • Pengyu Yuan is a Ph.D. student at the University of Houston