"Deep Learning-based MRI/Optical Registration for Neurosurgery"

Context
For the resection of brain tumors, it is crucial to precisely and robustly localize different cortical areas to avoid any postoperative sequelae. To achieve this, we need to register a preoperative 3D MRI with a 2D RGB image acquired during surgery. The goal of this project is to develop a method that performs this task in a robust and accurate manner. We will rely on deep learning approaches, which have demonstrated their ability to produce robust algorithms. 2D/3D registration is inherently challenging. Additional difficulties in this project arise from the limited amount of available data for training and the challenge of obtaining precise annotations for supervision.
This PhD will be conducted as part of a collaboration between CREATIS, HCL, and INRIA.

Challenges
The registration problem at the core of this thesis is particularly difficult due to the heterogeneity of modalities (3D MRI vs. 2D hyperspectral or RGB optical images). It will be necessary to align data with radically different resolutions and contrasts, and some structures may be visible in one image but not the other.
Clinical databases are limited in size and diversity (few aligned MRI/optical image pairs). The anatomical and pathological variability between patients makes it difficult to generalize models.
The proposed method is intended to be used in the operating room during surgery. It must meet specific constraints, particularly in terms of robustness and computational time.

Methodological Approach
From a methodological standpoint, the PhD will involve work in:

  • Representation learning
  • Learning with limited data
  • Semi/weakly supervised learning
  • Domain-specific data augmentation
  • 3D geometry consideration (computer vision)
  • Integration of physical constraints

Candidate Profile
We are seeking candidates with:

A Master’s degree or Engineering degree in signal processing, machine learning/AI, or applied mathematics.
Programming skills in Python and experience with deep learning libraries.
An interest in the biomedical field.

Bibliography
 

[1] Charly Caredda, Eric Van-Reeth, Michaël Sdika, Fernand Fort, Jacques Guyotat, Fabien C. Schneider, Thiébaud Picart, Bruno Montcel, Registration of Intraoperative Optical Imaging with Preoperative T1-Weighted MRI, submitted to IEEE Transaction on Biomedical Engineering

[2] Slim Hachicha, Célia Le, Valentine Wargnier-Dauchelle, Michaël Sdika. Robust Unsupervised Image to Template Registration Without Image Similarity Loss. Medical Image Learning with Limited and Noisy Data, Second International Workshop, MILLanD 2023, Held in Conjunction with MICCAI 2023, Vancouver, Proceedings, Oct 2023, Vancouver, Canada. https://hal.science/hal-04183379v1