Image Registration with Deep Learning
Recalage d'Images Médicales par apprentissage profond
Master II internship - 2022
(version française dans le pdf joint )
Keywords: Medical image registration, Deep Learning
Scientific background
Image registration is a tool to align images, allowing to re-position, re-orientate or even to deform images to replace them in a common coordinate system. This is often an essential preliminary step for the study of brain pathology based on imaging.
When the transformation is affine, current registration tools give results that are often satisfactory but but fail in particular when:
- initialization is bad
- there are strong imaging artifacts
- a pathology severely affect the brain appearance
- only part of the brain is present in the image (cropped image)
Classical registration tools are often based on iterative mathematical optimization approaches
but more and more current methods are based on deep learning approaches [Boveiri].
Objectives
The objective of the internship is to set up and train a neural network allowing
perform linear registration of a brain image to a standard reference space. The main objective
is to have a very robust estimation but also that the network is lightweight, fast to infer, and also interpretable to be able detect failures.
We can investigate good ways to parameterize the transformation, the equivariant layers [Finzi], capsule network [Sabour, Lensen, Gu].
The linear registration network will be integrated into the brain pre-processing pipeline of the
MYRIAD team of CREATIS.
Data: several public brain imaging datasets involving different pathologies, acquisition protocols and
pathologies, acquisition protocols and modalities are already used in the team and will be used for the
internship. A robust data augmentation procedure will further improve the robustness of our method.
Partners
This work will be done in Lyon in collaboration between CREATIS and LIRIS and will be supervised by:
- M. Sdika, CREATIS, Myriad team
- C. Garcia, LIRIS, Imagine team
Candidate Profile
The recruited candidate will have a background in one of the following areas and good knowledge in the other two:
- Machine learning (deep learning)
- Image processing
- Applied mathematics
He/she should also have strong software development skills and be able to implement the proposed
to implement the proposed methods.
Please send your application with CV, cover letter, transcripts, letters of recommendation to michael.sdika[at]creatis.insa-lyon.fr and Christophe.garcia[at]insa-lyon.fr.
References
- Boveiri et al, Medical Image Registration Using Deep Neural Networks: A Comprehensive Review, https://arxiv.org/pdf/2002.03401.pdf
- Finzi et al. A Practical Method for Constructing Equivariant Multilayer Perceptrons for Arbitrary Matrix Groups, https://arxiv.org/pdf/2104.09459.pdf
- Sabour et al, Dynamic Routing Between Capsules, https://arxiv.org/abs/1710.09829
- Lenssen et al, Group Equivariant Capsule Networks, Neurips 2018
- Gu J. & Tresp V., Improving the Robustness of Capsule Networks to Image Affine Transformations, CVPR 2020.