MLBDNI: Postdoctoral Position in Machine-Learning Biomarker Discovery in Neuroimaging

The Marie S. Curie Postdoctoral Fellowship (MSCA-PF) programme is a highly prestigious renowned EU-funded scheme. It offers talented scientists a unique chance to set up 2-year research and training projects with the support of a supervising team. Besides providing an attractive grant, it represents a major opportunity to boost the career of promising researchers.

The LTSI - UMR 1099 MediCIS Team, INSERM / Université de Rennes, is thus looking for excellent postdoctoral researchers with an international profile to write a persuasive proposal to apply for a Marie S. Curie Postdoctoral Fellowship grant in 2024 (deadline of the EU call set on 11 September 2024). The topic and research team presented below have been identified in this regard.

Research field: Information Science and Engineering (ENG), Artificial Intelligence, Computer Vision, Neuroimaging

Keywords: Artificial Intelligence, Computer Vision, Neuroimaging

Research project description:

Context:

Recent machine learning techniques have be lauded for their capacity to detect subtle changes in the brain associated with various neurological disorders (Nenning & Lang, 2022). However, due to their “black-box” nature, existing machine learning models do not provide much information that can elucidate what these subtle changes are, how they vary from patient to patient, and how those variations contribute to differences in symptoms or disease sub-types. This is particularly important for MRI-based neuroimaging where a detailed understanding of precise biomarkers could lead to earlier diagnosis and more accurate disease monitoring. The goal of this project is therefore to develop frameworks to discover, extract, and validate the biomarkers used by modern machine-learning architectures operating on volumetric neuroimages. This research will involve not only the design and implementation of new deep learning-based architectures for neurological disease diagnosis, but also population-based neuroimage analysis and interpretable machine learning algorithm design.

Objectives:

The overall project has three board objectives which will be defined in the following paragraphs. These objectives are all interdisciplinary, requiring the development of expertise in machine learning, probabilistic graphical models and other statistical techniques, MRI physics, and neuro-imaging.

The first technical aspect of this project is to re-implement existing highperformance deep learning architectures for disease classification via volumetric neuroimage analysis or to develop new light-weight architectures that are equally high-performance. By developing high-performance architectures for disease detection, we can be assured of the relevance of the imaging biomarkers that must underlie said performance.

Once a high-performance architecture has been designed and validated for a particular disorder, the second objective is to develop rigorous analysis tools for biomarker discovery. Due to the inherent stochastically present in machine learning algorithms, one would expect different instantiations of the same architecture to rely on different, and potentially erroneous, features. Thus, large-scale ensemble training of these architectures will be necessary to ensure that erroneous or weak biomarkers that only appear in a small number of networks are removed in favour of those that are robustly and independently used across multiple instantiations. In addition, population neuroimaging techniques such as voxel-based analysis will ensure the same for patients: that only biomarkers that consistently appear in a large number of patients for a large number of networks will be extracted as meaningful (Estudillo Romero et al., 2022a,b). This will also allow for an initial comparison between the extracted biomarkers and those from more traditional biomarker discovery approaches (such as voxel-based analysis) without the use of deep learning.

The final aspect of this project is to integrate these biomarkers into interpretable machine learning algorithms (Rudin et al., 2022) which can be readily understood by clinical users as well as more easily analyzed mathematically. This last point will allow for these biomarkers to be used in disorder monitoring and early diagnosis where the goal of the machine learning algorithm is less concerned with the detection of the disorder but rather a robust quantification as to how much of the disorder is manifested prior to the arrival of its characteristic symptoms. This should take into account the multiple ways in which the same disorder can manifest across different patients, leading again to a necessarily non-linear approach. This approach will not only encourage an independent evaluation of these biomarkers outside of deep learning architectures but will also encourage a focus on utility and acceptability, both to clinicians and to patients.

Each objective will be instantiated for multiple different neurological, neurodegenerative, and psychological disorders, making use of patient datasets acquired at the Rennes University Hospital, international collaborations, and open databases. This multi-disorder, multi-centric approach will ensure that the developed algorithms will be easily generalized and could encourage more holistic, rather than single-disorder, approaches to neuroimage analysis. In addition, this project will have to simultaneously use multiple MRI contrasts in order to exploit different physiological parameters but also be able to disentangle what modality or specific physiological parameter leads to particular algorithmic decisions.

1. Nenning, K. H., & Langs, G. (2022). Machine learning in neuroimaging: fromresearch to clinical practice. Die Radiologie, 62(Suppl 1), 1-10.

2. Rudin, C., Chen, C., Chen, Z., Huang, H., Semenova, L., & Zhong, C. (2022).Interpretable machine learning: Fundamental principles and 10 grand challenges.Statistic Surveys, 16, 1-85.

3. Estudillo Romero, A., Haegelen, C., Jannin, P., & Baxter, J. S. H. (2022a). Voxel‐based diktiometry: Combining convolutional neural networks with voxel‐basedanalysis and its application in diffusion tensor imaging for Parkinson's disease.Human Brain Mapping, 43(16), 4835-4851.

4. Estudillo Romero, A., Haegelen, C., Jannin, P., & Baxter, J. S. H. (2022b). Diffusiontensor imaging biomarkers for Parkinson’s disease symptomatology. MICCAI 2022Workshop on Medical Image Assisted Blomarkers' Discovery, 134-142.

Supervisors

The Postdoctoral Fellow will be co-supervised by John S.H. Baxter and Pierre Jannin from the MediCIS research group which is part of both Inserm (LTSI, UMR 1099) and the University of Rennes.

Pierre Jannin is an Inserm Research Director at the Faculty of Medicine of the University of Rennes (France) and director of the MediCIS research group. He has more than 30 years of experience in designing and developing computer assisted surgery systems. His research topics include surgical data science, surgical robotics, image-guided surgery, augmented and virtual reality, modeling of surgical procedures and processes, study of surgical expertise, and surgical training. He authored or co-authored more than 150 peer-reviewed international journal papers. He was the President of the International Society of Computer Aided Surgery (ISCAS) from 2014 to 2018. He is the Editor-in-Chief of Computer Assisted Surgery (Taylor&Francis).

John S.H. Baxter is an Inserm Researcher affiliated with the University of Rennes (France) studying the intersection of machine learning and human computer interaction for medical imaging. Despite only obtaining his Ph.D. in 2017, John already has 20 first- or last- authored peer-reviewed papers in international journals and is part of the scientific committees of the annual conference on Computer Assisted Radiology and Surgery (CARS) and the SPIE Medical Imaging Conference on Image-Guided Procedures, Robotic Interventions and Modeling. He is also a member of the MICCAI Society Office which organises the annual conference on Medical Image Computing and Computer Assisted Interventions (MICCAI).

MediCIS webpage: https://medicis.univ-rennes1.fr/

Pierre Jannin profile: https://scholar.google.com/citations?user=yr_qKA0AAAAJ

John S.H. Baxter profile: https://scholar.google.com/citations?user=zB8zUmYAAAAJ

Department

The Laboratory of Signal and Image Processing (LTSI), is an INSERM laboratory at the University of Rennes with about 150 researchers dedicated on Biomedical Engineering research.

The MediCIS team, located at the medical faculty, is part of the LTSI and focuses on the study of modeling aspects of medical knowledge and skills acquisition through virtual reality, machine learning, medical image processing, and surgical data science all with the participation of clinicians from the Rennes University Hospital. This research team published pioneering work in surgical data science, machine learning in functional neurosurgery, surgical workflow analysis with surgical procedure models, and ontologies for medical imaging and surgery.

https://medicis.univ-rennes1.fr/