Keynotes

Nina Miolane

Assistant Professor, UC Santa Barbara

Shape Learning in Biomedical Imaging

Advances in biomedical imaging techniques have enabled us to observe biological structures in living organisms: proteins, cells, organs. Each of these structures adopts a variety of shapes, depending on its physiological function, or healthy versus pathological state. Shape data analysis is thus essential to understand the fundamental mechanisms of life. This leads to the question: how can we learn quantitative descriptions of shape variability from images? This talk describes geometric (deep) learning for biomedical shape data analysis and introduces its geometric computational building blocks via the open-source package Geomstats.

Biography

Nina Miolane received her M.S. in Mathematics from Ecole Polytechnique (France) & Imperial College (UK), and her Ph.D. in Computer Science from INRIA (France) in collaboration with Stanford University. After her studies, Nina spent two years at Stanford University in Statistics as a postdoctoral fellow, and worked as a deep learning software engineer in Silicon Valley. At UCSB, Nina directs the BioShape Lab, which investigates (i) how the shapes of proteins, cells, and organs relate to their biological functions, (ii) how abnormal shape changes correlate with pathologies, and (iii) how these findings can help design new automatic diagnosis tools. Her team also co-develops the open-source Geomstats library, a software that provides methods at the intersection of geometry and machine learning, to compute with geometric data such as biological shape data. Nina was the recipient of the L'Oréal-Unesco for Women in Science Award.

Jessica Zhang

Professor, Carnegie Mellon University

Machine Learning Enhanced Simulation and PDE-Constrained Optimization for Material Transport Control in Neurons

The intracellular transport process plays an important role in delivering essential materials throughout branched geometries of neurons for their survival and function. Many neurodegenerative diseases have been associated with the disruption of transport. Therefore, it is essential to study how neurons control the transport process to localize materials to necessary locations. First, we develop an isogeometric analysis (IGA) based platform for material transport simulation in neurite networks. We model the transport process by reaction-diffusion-transport equations and represent geometry of the networks using truncated hierarchical tricubic B-splines (THB-spline3D). We solve the Navier-Stokes equations to obtain the velocity field of material transport in the networks. We then solve the transport equations using the streamline upwind/Petrov-Galerkin (SU/PG) method. Next, we develop a novel optimization model to simulate the traffic regulation mechanism of material transport in neurons. The transport is controlled to avoid traffic jam of materials by minimizing a pre-defined objective function. The optimization subjects to a set of partial differential equation (PDE) constraints that describe the material transport process based on a macroscopic molecular-motor-assisted transport model of intracellular particles. Different simulation parameters are used to introduce traffic jams and study how neurons handle the transport issue. Our model effectively simulates the material transport process in healthy neurons and explains the formation of a traffic jam caused by reduced number of microtubules (MTs) and MT swirls in abnormal neurons. To enable fast prediction of the transport process within complex neurite networks, we develop a Graph Neural Networks (GNN) based model to learn the material transport mechanism from simulation data. In this study, we build the graph representation of the neuron by decomposing the neuron geometry into two basic structures: pipe and bifurcation. Different GNN simulators are designed for these two basic structures to predict the spatiotemporal concentration distribution given input simulation parameters and boundary conditions. In particular, we add the residual term from PDEs to instruct the model to learn the physics behind the simulation data. To recover the neurite network, a GNN-based assembly model is used to combine all the pipes and bifurcations following the graph representation. The loss function of the assembly model is designed to impose consistent concentration results on the interface between pipe and bifurcation. Through machine learning, we can quickly and accurately provide a prediction of complex material transport patterns including traffic jam and MT swirls.

Biography

Jessica Zhang is the George Tallman Ladd and Florence Barrett Ladd Professor of Mechanical Engineering at Carnegie Mellon University with a courtesy appointment in Biomedical Engineering. She received her B.Eng. in Automotive Engineering, and M.Eng. in Engineering Mechanics from Tsinghua University, China; and M.Eng. in Aerospace Engineering and Engineering Mechanics and Ph.D. in Computational Engineering and Sciences from Institute for Computational Engineering and Sciences (now Oden Institute), The University of Texas at Austin. Her research interests include computational geometry, isogeometric analysis, finite element method, data-driven simulation, image processing, and their applications in computational biomedicine, materials science and engineering. Zhang has co-authored over 200 publications in peer-reviewed journals and conference proceedings and received several Best Paper Awards. She published a book entitled "Geometric Modeling and Mesh Generation from Scanned Images" with CRC Press, Taylor & Francis Group. Zhang is the recipient of Simons Visiting Professorship from Mathematisches Forschungsinstitut Oberwolfach of Germany, US Presidential Early Career Award for Scientists and Engineers, NSF CAREER Award, Office of Naval Research Young Investigator Award, and USACM Gallagher Young Investigator Award. At CMU, she received David P. Casasent Outstanding Research Award, George Tallman Ladd and Florence Barrett Ladd Professorship, Clarence H. Adamson Career Faculty Fellow in Mechanical Engineering, Donald L. & Rhonda Struminger Faculty Fellow, and George Tallman Ladd Research Award. She is a Fellow of AIMBE, ASME, SMA, USACM and ELATES at Drexel. She is the Editor-in-Chief of Engineering with Computers.

Tolga Birdal

Assistant Professor, Imperial College London

Rigid & Non-Rigid Multi-Way Point Cloud Matching via Late Fusion

Correspondences fuel a variety of applications from texture-transfer to structure from motion. However, simultaneous registration or alignment of multiple, rigid, articulated or non-rigid partial point clouds is a notoriously difficult challenge in 3D computer vision. With the advances in 3D sensing, solving this problem becomes even more crucial than ever as the observations for visual perception hardly ever come as a single image or scan. In this talk, I will present an unfinished quest in pursuit of generalizable, robust, scalable and flexible methods, designed to solve this problem. The talk is composed of two sections diving into (i) MultiBodySync, specialized in multi-body & articulated generalizable 3D motion segmentation as well as estimation, and (ii) SyNoRim, aiming at jointly matching multiple non-rigid shapes by relating learned functions defined on the point clouds. Both of these methods utilize a family of recently matured graph optimization techniques called synchronization as differentiable modules to ensure multi-scan / multi-view consistency in the late stages of deep architectures. Our methods can work on a diverse set of datasets and are general in the sense that they can solve a larger class of problems than the existing methods. The relevant publications include:

MultiBodySync: https://arxiv.org/abs/2101.06605 [CVPR’21 Oral]

SyNoRiM: https://arxiv.org/abs/2111.12878 [T-PAMI]

Biography

Tolga Birdal is an assistant professor in the Department of Computing of Imperial College London. Previously, he was a senior Postdoctoral Research Fellow at Stanford University within the Geometric Computing Group of Prof. Leonidas Guibas. Tolga has defended his masters and Ph.D. theses at the Computer Vision Group under Chair for Computer Aided Medical Procedures, Technical University of Munich led by Prof. Nassir Navab. He was also a Doktorand at Siemens AG under supervision of Dr. Slobodan Ilic working on “Geometric Methods for 3D Reconstruction from Large Point Clouds”. His current foci of interest involve geometric machine learning and 3D computer vision. More theoretical work is aimed at investigating and interrogating limits in geometric computing and non-Euclidean inference as well as principles of deep learning. Tolga has several publications at the well-respected venues such as NeurIPS, CVPR, ICCV, ECCV, T-PAMI, ICRA, IROS, ICASSP and 3DV. Aside from his academic life, Tolga has co-founded multiple companies including Befunky, a widely used web-based image editing platform.