Learn2Reg 2019: Tutorial on Deep Learning in Medical Image Registration

Overview

Medical image registration has been a cornerstone in the research fields of medical image computing and computer assisted intervention, responsible for many clinical applications. Whilst machine learning methods have long been important in developing pairwise algorithms, recently proposed deep-learning-based frameworks directly infer displacement fields without iterative optimisation for unseen image pairs, using neural networks trained from large population data. These novel approaches promise to tackle several most challenging aspects previously faced by classical pairwise methods, such as high computational cost, robustness for generalisation and lack of inter-modality similarity measures. Output from several international research groups working in this area include award-winning conference presentations, high-impact journal publications, well-received open-source implementations and industrial-partnered translational projects, generating significant interests to all levels of world-wide researchers. Accessing to the experience and expertise in this inherently multidisciplinary topic can be beneficial to many in our community, especially for the next generation of young scientists, engineers and clinicians who often have only been exposed to a subset of these methodologies and applications. We propose to organise a tutorial including both theoretical and practical sessions, inviting expert lectures and tutoring coding for real-world examples. Six lectures cover topics from basic methodologies to advanced research directions, with two hands-on sessions guiding participants to understand and implement published algorithms using clinical imaging data. We aim to provide an opportunity for the participants to bridge the gap between expertises in medical image registration and deep learning, as well as to start a forum to discuss know-hows, challenges and future opportunities in this area.

Tutorials are available on github.

Program

Morning Session

• 9:00 - 9:30 --- Tom Vercauteren
  Introduction to Medical Image Registration
• 10:00 - 11:00 --- Adrian V. Dalca
  Unsupervised and Self-supervised Methods
  Hands-on session: unsupervised registration based tutorial
• 11:00-12:00 --- Yipeng Hu
  Supervised and Weakly-supervised Methods
  Hands-on session: label-supervised based tutorial

Afternoon Session

• 14:00- 14:30 --- Julia Schnabel
  Opportunities and Challenges for Data Fusion
• 14:30- 15:00 --- Bob de Vos
  Unsupervised deep learning image registration: Beyond the cranial vault
• 15:00-16:00 --- Mattias Heinrich
  Discrete Deep Learning Registration
  Hands-on session: Discrete Deep Learning Registration.
• 16:30-17:30 Panel Discussion and Discussion about Hand-on Sessions

Details

Outline

Recent machine learning methods based on deep neural networks have seen a growing interest in tackling a number challenges in medical image registration, such as high computational cost for volumetric data and lack of adequate similarity measures between multimodal images [de Vos et al, Hu et al, Balakrishnan et al, Blendowski & Heinrich, Eppenhof & Pluim, Krebs et al, Cao et al.]. Other interesting work include learning new similarity measure functions, incorporating prior knowledge such as diffeomorphism and estimating registration uncertainties [Cheng et al, Simonovsky et al].

Objectives

Six lectures are planned on topics from classical image registration methodology to practical algorithms using deep-learning, including an introduction to image registration, unsupervised and supervised learning methods, similarity measure learning, and an outlook to opportunities and challenges. These will be accompanied by a set of open-source, interactive tutorials. The expected outcome include a) a basic understanding of medical image registration; b) an overview of modern deep-learning based methods; c) hands-on experience in adopting and adapting open-source code to train and deploy registration networks for real-world clinical problems.

Interests

A strong interest in deep-learning applied on image registration can be demonstrated by the number of papers recently published in venues such as MICCAI, MedIA and IEEE-TMI related to this topic. For example, the paper [de Vos et al] addressing this topic published in 2017 won the workshop’s best-paper prize and has been well received. The open-source code maintained by two of our organisers has attracted significant interest, VoxelMorph and LabelReg, among other closely-relevant emerging topics, such as AIRLab. The potential benefits for easing the accessibility of these new methods to wider practitioners in MICCAI community is considerable. For instance, an automatic advantage of the learning-based registration methods comes from highly efficient inference - an enabling aspect for many large-scale analysis and interventional applications. We would also like to emphasize that, due to the cross-topical nature and high relevance to many clinical applications and researchers, placing this proposed tutorial in the “pre-conference” slots at the first day of MICCAI will maximise its benefits by equipping those interested to gain knowledge and relevant topics during the main conference - in which we are expecting significantly more submissions under this topic in 2019.