Artificial Intelligence for medical image analysis : application to functional ultrasound imaging

functional UltraSound imaging (fUSi) is a technology developed in our laboratory (Urban et al., 2015; Macé et al., 2018) allowing real-time visualization of brain activity with a high spatiotemporal resolution (100µm3 voxel size, 100ms). It has been widely validated in preclinical research but fUSi is currently limited to imaging of a small part of rodent’s brain. To extend fUSi capabilities, we recently developed 3D-fUSi for volumetric imaging of entire mouse brain dynamics generating complex datasets (500k voxels/image) at high imaging rate (10Hz). We are now aiming to perform 3D-fUSi data analysis in real-time including both i) automated segmentation/parcellation of images in regions of interest based on the brain atlas and ii) resting and dynamic functional data analyses. Most of data processing algorithms are not suited for real-time analysis of high-dimensional imaging data and require manual intervention. We aim at developing new Artificial Intelligence-based methods offering quick and automated analysis of 2D/3D-fUSi data. Our strategy is to perform preprocessing (filtering, denoising, features extraction, …) and potentially data augmentation using deep generative models - such as generative adversarial networks - or autoencoders. Then, we will use Convolutional Neural Networks and/or Recurrent Neural Networks suitable for accurate segmentation with a limited number of training samples. During this PhD, we will investigate the most relevant deep learning algorithms and potentially combine them with existing segmentation tools to achieve the best speed/accuracy trade-off. We will explore machine learning methods to automatically detect and extract complex brain activity patterns linked to different behavioral tasks using unsupervised learning, probabilistic graphical models (as random Markov fields), probabilistic mixtures and/or Bayesian modelling. A secondary objective is to evaluate the benefit of high-performance computing and software optimization for real-time data analysis in collaboration with imec Exascience team.