Le Forum de statistique organisé par Unisanté et le Département de Biologie Computationnelle (DBC) de l’UNIL a lieu huit fois par an. Toutes sortes de thématiques de statistiques liés à la recherche scientifique, que ce soit d’un point de vue théorique ou appliqué sont abordées lors de séminaires, donnés en anglais ou en français.
Le forum est ouvert aux chercheurs et chercheuses de toutes institutions. Toutes autres personnes intéressées sont également les bienvenues.
Les séminaires pourront être suivis en visioconférence.
Graph-based learning in medical imaging and imaging genetics
Graphs are particularly well-suited for modelling relationships between parts of a system. This includes organs such as the brain, where graph representations are ubiquitous and offer an expressive language to model spatial, structural, and functional relationship from data obtained in medical imaging. Graphs enable substantial compression and smoothing of imaging data, which helps build predictive models from high-dimensional volumetric time series such as found in functional magnetic resonance imaging (brain) or CINE imaging (heart).
In this overview talk, I will first discuss estimation of graph edges, in the first instance more specifically the estimation of intrinsic correlations between brain regions. Drawing links with work in familial data and geostatistics, I will present some ongoing work about correlation estimators using replicates for 4D brain imaging data (3D + time), with a focus on robustness to local versus global noise. I will also show related ongoing work using image features to estimate edges in multiplex (multilayer) graphs for 4D cardiac imaging. This can represent cardiac structure and function, including morphology and motion.
Once a graph is obtained from the images of the organ, multiple approaches can be used to relate it to the clinical score or phenotype of interest. I will first briefly discuss network science approaches and show how they can be used to inform our understanding of organ function in health and disease, and then focus on machine learning approaches for graph data. I will discuss more specifically embedding approaches and graph neural networks, which have shown outstanding performance even in the limited data regime.
Finally, I will discuss how graphs computed from medical imaging data can be used in imaging genetics as an endophenotype representation and target for association studies, including genetic variation and transcriptomics. Increasingly, large databases suitable for imaging genetics studies are being collected (most notably UK Biobank), and I will briefly present our work in the area of heart failure subtyping using heart imaging and genetic data.
Jonas Richiardi is a Principal Investigator and Senior Lecturer at the Department of Radiology, Lausanne University Hospital, Switzerland, and heads the Translational Machine Learning Laboratory there. Previously, he was the Clinical Research Lead at Siemens Healthcare's Advanced Clinical Imaging Technology. Prior to this he was a Marie Curie fellow at Stanford University (Neurology) and the University of Geneva (Fundamental Neurosciences), and post-doctoral researcher in the Medical Image Processing Lab, a joint position between the Ecole Polytechnique Fédérale de Lausanne (EPFL), Institute of Bioengineering, and the University of Geneva's Department of Radiology and Medical Informatics. He obtained his Ph.D. in signal processing and pattern recognition in 2007 at EPFL in the Signal Processing Institute (Laboratory of IDIAP), and his M.Phil. in speech and language processing and modelling from the university of Cambridge (Computer Science and Engineering Departments).
His research interests include modelling and inference for complex multimodal biological data, in particular magnetic resonance imaging data and its combination with genetic, transcriptomic, metabolomic, and proteomic data. He focuses on graph-based statistical learning approaches, where all data is first represented as a graph, and machine learning approaches are applied to form prediction with graphs. Applications are early diagnosis, prognosis, and treatment response prediction for individual subjects, for diseases ranging from Alzheimer's disease to stroke, multiple sclerosis, and heart failure. An ongoing effort is to develop these techniques so that they can be applied to messy, hospital-scale data, which goes together with an important effort to improve the data science infrastructure at the Department of Radiology.
Retrouvez les enregistrements vidéo des séminaires passés ici.
Zoltan Kutalik
Responsable de l'Unité statistiques génétiques à Unisanté
E-mail
021 314 67 50