PhD position 08 – MSCA COFUND, AI4theSciences (PSL, France) - “Physically Informed Machine Leaning...
Universite PSL
Sophia Antipolis, France
il y a 4j
source : Euraxess

Artificial intelligence for the Sciences (AI4theSciences) is an innovative, interdisciplinary and intersectoral PhD programme, led by Université Paris Sciences et Lettres and co-funded by the European Commission.

Supported by the European innovation and research programme Horizon 2020-Marie Sklodowska-Curie Actions, AI4theSciences is uniquely shaped to train a new generation of researchers at the highest academic level in their main discipline (Physics, Engineering, Biology, Human and Social Sciences) and master the latest technologies in Artificial Intelligence and Machine Learning which apply in their own field.

26 doctoral students will join the PSL university's doctoral schools in 2 academic cohorts to carry out work on subjects suggested and defined by PSL's scientific community.

The 2020 call will offer up to 15 PhD positions on 24 PhD research projects. The candidates will be recruited through HR processes of high standard, based on transparency, equal opportunities and excellence.

Description of the PhD subject : Physically Informed Machine Leaning for controlling unruptured intracranial aneurysms

Context - Motivation

Cerebral aneurysms are balloonings of blood vessels in the brain and are thought to be present in 1 in 30 adults . Most of these aneurysms may never rupture, but when they do it is catastrophic : roughly half of patients die and of those that survive, roughly one-third will be permanently disabled.

Because brain imaging is being used widely as part of routine clinical workups, unruptured aneurysms are being detected more frequently.

This poses a dilemma for the doctor and patient since the risk associated with treating the aneurysm can often be higher than the risk of rupture.

Currently, doctors base their decisions on the size, shape, and location of the aneurysm, as well as family history and other risk factors such as smoking and hypertension.

But these together are not strong predictors of rupture risk.

Research over the past decades has demonstrated that the hemodynamic forces exerted by the complex blood flow patterns within aneurysms may provide important clues as to the state of the aneurysm wall.

Among several hemodynamic parameters that are discussed as key factors in the initiation, development, or rupture of intracranial aneurysms, one of the most studied parameters is the wall shear stress (WSS, the frictional forces exerted by flowing blood).

High or low local values of the WSS and non-uniform distribution of instability are negative conditions for the development of an aneurysm.

Indeed low WSS may lead to spatial disorganization of endothelial cells and a disregulation of antioxidant and anti-inflammatory mediators result in arterial wall remodeling.

Consensually, high WSS may lead to the initiation of aneurysm formation, but its influence on the growth and rupture is largely unknown.

Therefore, the availability of a simulation tool that can predict the aneurysm hemodynamics parameters on an individual basis will be extremely useful, either to develop new devices, or to support treatment decisions.

Owing to the absence of precise in vivo measurement tools, computational techniques offer new capabilities in the healthcare provision for cerebral aneurysms 1 .

Significant studies have been led by modeling the blood flow behavior using computational fluid dynamics (CFD) and Fluid-Structure interaction (FSI) simulations .

CFD-FSI is a powerful tool not only used to design cars and airplanes but also in medical applications.

Nevertheless, these simulations require many assumptions, including the challenging patient-specific modeling of the vascular system, and require significant computational resources.

Recent advances in the development of Deep Reinforcement Learning (DRL) algorithms have led to the rise of Deep Neural Networks (DNN), powerful tools capable of leveraging the ever-increasing volume of numerical and experimental data generated for research and engineering purposes into novel insight and actionable information.

State-of-the-art DRL techniques have proven fruitful for various applications, from solving computer vision problems to achieving super-human level in complex games, and have led breakthroughs in the optimal control of complex dynamic systems.

Given the ability of DNNs to handle large scale, non-linear systems, it is only natural to attempt to use them to tackle similarly the state-space models resulting from the high-dimensional discretization of partial differential equations Navier-Stokes equations) 2 .

Following these striking achievements, DRL has consistently spread to fluid mechanics, and has led to a handful of seminal, high-potential techniques addressing flow control and optimization problems.

Scientific Objectives, Methodology & Expected results

The proposed research aims then to bridge this gap between high fidelity CFD for the prediction of the aneurysm hemodynamics with the DRL for flow control and optimization which is importantly useful to reduce or to exclude the aneurysm from circulation 3 by employing the adequate stent for flow diversion.

Therefore, the size, the location and the shape of such stent will be deduced by the proposed CFD-DRL framework. The deployment and the analysis of fifty consecutives patients with unruptured intracranial aneurysms, provided by the Neuroradilogy and Intervention Institute from the University Hospital-LMU Munich, will be included in this study and they will serve for our data sets, training, tests and validations.

These geometries have been reconstructed from MR angiography images acquired with 3D time-of-flight (TOF) technique. The resulting images will be subsequently segmented in order to extract the unruptured intracranial aneurysms.

Different segmentation techniques will be explored in this purpose, based upon a novel approach recently developed for computing superpixel segmentations 4 and for merging superpixels together in order to get an actual segmentation of the aneurysms.

Supervised approaches including convolutional neural networks will also be considered. This will definitely form the required data set for test and validations.

Finally, a qualitative comparison with the patient-specific digital subtraction angiography (DSA) will be performed. Predictions based on deep learning based CFD simulations can predict rupture risk of the aneurysm, anticipate an optimal hemodynamic pattern before the clinical procedure, consequently decreasing the complications, improving life quality and expectancy of patients, and ultimately helping saving lives.

The proposed work is highly multidisciplinary and the algorithms developed as a part of this research can be quickly adopted to a wide range of engineering and medical applications.

The research will also greatly encourage candidates to pursue careers in interdisciplinary areas that bridge the biological, mathematical, and computational sciences.

Main references :

1 Y. Qian, H. Takao, M. Umezu and Y. Murayama Risk Analysis of Unruptured Aneurysms Using Computational Fluid Dynamics Technology : Preliminary Results, American Journal of Neuroradiology November 2011, 32 (10) 1948-1955

2 Optimization and passive flow control using single-step deep reinforcement learning, H. Ghraieb, J. Viquerat, A. Larcher, P.

Meliga, E. Hachem, Submitted to Physical Review Fluids,

3 Kim, Heung Cheol et al. Machine Learning

Signaler cette offre d'emploi

Thank you for reporting this job!

Your feedback will help us improve the quality of our services.

Mon email
En cliquant sur « Continuer », je consens au traitement de mes données et à recevoir des alertes email, tel que détaillé dans la Politique de confidentialité de neuvoo. Je peux retirer mon consentement ou me désinscrire à tout moment.
Formulaire de candidature