PhD scholarship Deep learning image analysis and computational simulation for the analysis of perinatal brain development

During the early phases of foetal development, the human brain undergoes a remarkable process of surface expansion and folding, resulting in the formation of outwardly protruding folds called gyri and inwardly concave folds known as sulci. Research findings indicate that several neurological disorders such as schizophrenia, autism, and lissencephaly, are associated with abnormal folding patterns. These findings underscore the significance of studying cortical folding and its potential implications for understanding and addressing these conditions.

In the last decades, computational modelling has emerged as a promising tool to understand the process of cortical folding, with numerous models based on different continuum mechanics theories. Integrating patient-specific longitudinal data is crucial for accurate predictions of the models with parameter optimization processes. Magnetic resonance imaging (MRI) has proven invaluable in studying brain development, but acquiring high-quality images from foetuses is very challenging due to the presence of movement-induced artifacts and the small size of the structures to capture.

During the last years, researchers from the BCN MedTech (https://www.upf.edu/web/bcn-medtech) research unit at Universitat Pompeu Fabra (UPF) have developed advanced computational tools to better understand brain development, in collaboration with several national and international clinical and academic partners. For instance, we recently proposed [1] a computational pipeline using foetal and neonatal MRI to generate patient-specific meshes to perform personalized simulations of brain growth and folding from the foetal stage, deriving multiple local and global metrics to characterise brain data and simulations. Furthermore, advanced processing pipelines have deep learning (DL) algorithms have been developed [2] to register foetal and neonatal MRI data, finding the displacement field between the two stages to analyse regional deformations. New methodological approaches for characterising longitudinal data, based on normalised flows, have also been proposed [3]. Research leading to these scientific contributions was possible due to the long-standing collaboration between UPF and the BCN Natal unit at Hospital Clínic and Hospital Sant Joan de Déu in Barcelona, particuarly with Dr. Elisenda Eixarch, world-wide expert on neurodevelopment and foetal therapies. Dr. Eixarch’s research group is a reference centre on foetal MRI, having access to unique clinical databases that are key to jointly advance the knowledge on brain development. Dr. Eixarch is also the principal investigator of a European project, MULTI-FACT, where, together with UPF and other international partners, are stablishing good practices for processing foetal MRI data with advanced machine learning algorithms, including the exploration of federated learning, and computational models to better understand the main characteristics leading to abnormal brain development.

Despite the progress achieved both in computational modelling and deep learning techniques in recent years, it is only now that researchers are combining the strengths from each technology, e.g., for accelerating simulations. For brain development modelling, having fast simulations would help to create more realistic simulations by embedding them into parameter optimization processes. At the time being, this is not possible due to the lack of large longitudinal datasets (with MRI data at foetal and neonatal stages from the same case) and the extensive computational time required for brain mechanical simulations (up to 35 hours from GA 26 to 40). However, having access to the unique clinical datasets from our clinical collaborators, including the dHCP data (http://www.developingconnectome.org/project), and the tools developed at UPF, including brain mechanical models and DL-based MRI analysis, create the ideal conditions for building the next generation of DL-based tools for characterising normal and abnormal brain development. 

[1] Alenyà, M., Wang, X., Lefèvre, J., Auzias, G., Fouquet, B., Eixarch, E., Rousseau, F., Camara, O.: Computational pipeline for the generation and validation of patient-specific mechanical models of brain development. Brain Multiphysics 3 (2022).

[2] Comte V., Alenyà M., Urru A., Nakaki A., Crovetto F., Camara O., Eixarch E., Crispi F., Piella G., Ceresa M., González Ballester M.A. Unsupervised segmentation of fetal brain MR images using multi-atlas segmentation and cascaded registration. International Journal of Computer Assisted Radiology and Surgery 18(S1).

[3] Masias M., González Ballester M.A., Piella G. Predicting structural brain trajectories with discrete optimal transport normalizing flows. Medical Imaging Meets Neurips (Med-Neurips), New Orleans, USA, pp. 99.1-4, 2022.

PhD project

In collaboration with several junior/senior researchers at UPF and external collaborators, the PhD candidate will work on the development of deep-learning-based algorithms for the analysis of longitudinal MRI data and the computational modelling of brain development. Initially, fast image registration techniques will be applied between foetal and neonatal imaging data to characterise deformation along time. The resulting deformation field would be introduced to guide the optimisation of brain mechanical models, thus improving the realism of the resulting simulations. Fast surrogate models of brain mechanics will be investigated, based on the current state-of-the-art techniques (e.g., PINNs, DeepONets, etc.), some of which have already tested at UPF. The PhD will be directly supervised by Prof. Miguel A. González Ballester and Prof. Oscar Camara from BCN MedTech at UPF.

Profile of the candidate

We are looking for highly motivated young researchers with a MSc degree (or equivalent) in Biomedical Engineering, Data Science, Physics, Mechanical Engineering, Applied Mathematics, Computational Science, or related disciplines, willing to study and do research at the leading edge of biomedical engineering. Experience in computer sciences and having proven programming skills would be of importance. High motivation is the only essential pre-requisite. Nevertheless, candidates already familiar with deep learning methods for medical image processing (e.g., nn-U-net, VoxelMorph, etc.), and for building fast surrogates of computational physics-based models (e.g., PINN, DeepONets, POD), would have a faster start of the project.

Candidates must have excellent teamwork and communication skills and be enthusiastic about collaborating with a diverse range of international partners. We expect them to be fluent in oral and writing English. Interest in clinical translation is essential since meetings with clinicians will regularly take place.

More information on the requirements for a PhD position at the Universitat Pompeu Fabra can be found on https://www.upf.edu/web/etic/doctorat and http://www.upf.edu/doctorats/en.

Conditions

This position includes a teaching commitment load of 45 hours per academic year.

Starting date (planned): Oct/Nov 2023

Gross yearly salary: 20.200€ (1st and 2nd year), 21.043€ (3rd year), 24.204€ (4th year).

Application process

Please send your CV and motivation letter, along with any additional relevant material, to Prof. Miguel A. González Ballester <ma.gonzalez@upf.edu> and Prof. Oscar Camara <oscar.camara@upf.edu>.

Deadline: 31st July 2023.