DREAM 2020: Disentangled Representations for Efficient Algorithms for Medical data

A MICCAI 2020 (Peru) Tutorial by Sotirios A. Tsaftaris and Alison Q. O’Neil.

Outline

Disentangled representation learning has been proposed as an approach to learning general representations. This can be done in the absence of annotations, or with limited annotation. A good general representation can be readily fine-tuned for new target tasks using modest amounts of data. This alleviation of the data and annotation requirements offers tantalising prospects for tractable and affordable healthcare. Finally, disentangled representations can offer model explainability, increasing their suitability for real-world deployment.

In this half-day tutorial, a satellite event in conjunction with MICCAI 2020 (Peru), we will offer an overview of representation learning and disentangled representation learning and criteria, and discuss applications in medical imaging and the wider spectrum of EHR data. We will conclude with open ended challenges.

Program

Virtual Format: On the MICCAI platform we have uploaded several videos that cover the material of the tutorial. Please send us your questions before the session via email or via the platform. On the day of the tutorial, we will be offering an abridged version of the pre-recorded video material currently available on the platform LIVE to further develop audience interaction and ensure that everyone in the audience is on a `level’ playing field. We will divide the tutorial in sessions, keep each session short, and have live Q&A immediately after each session. These presentations do not replace the videos available already but offer a shorter summary. If you have seen the videos already and have emailed us questions before or submitted questions in the platform chat and we did not reply we will aim to reply to them live here during the Q&A. If we run out of time, please do reach out to us offline!

All times are UTC (for BST (current UK time), add +1). The full schedule is visible below.

Learning Objectives

• Understand representation space and why (in)variance matters
• Understand the theories of information bottleneck and compositionality
• Learn the different objectives in achieving disentanglement
• Appreciate the inductive biases introduced by network design choices
• Appreciate when disentanglement is useful in practice

Motivation

Imagine that we want to develop a system that localises the heart in MRI and CT images. This system will need to be robust to any changes in imaging, the scanner, noise, and critically anatomical and pathological variation. The current paradigm with deep learning is that we must present to the system as many examples as possible to make it robust and learn what is nuisance (e.g. noise and imaging differences) as opposed to what matters (e.g. the location of the heart). [A similar argument can be made for organs in a CAI setting.]

Clearly this is not sustainable and leads to poor performance. Disentangled learning can help because it allows us to learn latent factors that can describe what we see in the data. In the example on the right, we can imagine a latent factor that one learns a change in pose (or in the patient for the cardiac example), and another factor for the change in a car’s colour (or a different scanner). Surprisingly we do not always need annotated data to achieve this. Moreover, it has been shown that disentangled representations are privacy preserving; can offer explainability and interpretability; and can generalise to new tasks (meta-learning) and to new data sources with less effort.

We should (re)appreciate that machine learning (ML) is not simply a functional mapping between input and output, but one that maps input data to manifolds and then decisions/tasks. This return to principled design will allow us to evolve and propose efficient solutions that are robust to new applications and shorten the path to clinical translation.

Learning suitable representations is key in ML. In fact, the ML community has the dedicated International Conference on Learning Representations (ICLR). Disentangled learning is considered a hot area: a dedicated challenge run at NeurIPS 2019. According to Google Scholar trends, the number of papers in major ML conferences doubles yearly . Scientific American, a prestigious periodical in popular science, had a description of disentangled representation learning in May 2019 (Machine Learning Gets a Bit More Humanlike). Disentangled learning appeared in MICCAI 3 years ago in a couple of papers, but now has tripled in appearance. Thus, it is timely to formally introduce this important area of research in our community.

Teachers

Prof. Sotirios A Tsaftaris (Sotos) is Chair in Machine Learning and Computer Vision with the University of Edinburgh and is also the Canon Medical/Royal Academy of Engineering Research Chair in Healthcare AI. Sotos leads a group where several young researchers work in machine learning and computer vision. He obtained his PhD in 2006 from Northwestern University USA and has held several academic positions in USA, Italy and UK. Sotos’s research expertise is in representation learning.

Dr Alison Q O’Neil (Alison) is a Senior Scientist in the AI Research Team at Canon Medical Research Europe and Honorary Research Fellow at the University of Edinburgh. She obtained her EngD at Canon Medical Research Europe in affiliation with Heriot-Watt University, and now leads a team of scientists and research students working on machine learning techniques for industrial healthcare applications – including applications in medical imaging, natural language processing, and electronic health record (EHR) data. Alison will bring an industry perspective to the tutorial and also talk about disentanglement in other data forms (e.g. text).

Materials

See MICCAI platform.

Support

Generously supported by Canon Medical Research Europe, the Royal Academy of Engineering and the School of Engineering.