AMS Special Seminar Series | Mazdak Abulnaga

Location: Shaffer 3

When: February 26th at 1:30 p.m.

Title: Learning Biomedical Shape Mapping

Abstract: Machine learning is increasingly shaping scientific discovery and clinical decision-making in the life sciences. In many biomedical domains, biological function is affected by three-dimensional (3D) structure. For example, brain organization is tightly coupled to cortical folding patterns and geometry. Explicitly modeling the relevant 3D structure enables machine learning methods to extract more meaningful patterns and, in turn, deepen scientific understanding of biomedical function. 

In this talk, I will describe methods to model the physical structure of biomedical organs in machine learning. I will present new mapping methods that estimate a transformation from a volumetric source shape to a target domain. Producing such mappings is a critical step in scientific studies that require comparative analyses of biological structures and function. Current mapping methods do not model both volumetric and surface information, dramatically limiting their applicability in biomedical and physical domains. I will first present a technique to map highly irregular deformable organs, such as the placenta, to a canonical template for visualization and analysis.  Building on these concepts, I will describe a machine learning method that jointly maps the highly folded surface of the brain and the interior volume to enable rapid large-scale whole-brain neuroscientific analysis. Finally, I will then describe methods for fast and flexible construction of image templates, enabling population analyses with new imaging and data collections. I will conclude with a brief look at future work. I plan to develop unified representations of function and shape from biomedical systems, incorporating multimodal health and biological data across all scales from proteins to organs to enable novel studies in biology and medicine.

Bio: Mazdak Abulnaga is a postdoctoral researcher at the MIT Computer Science and Artificial Intelligence Lab and Harvard Medical School, working with Adrian Dalca, John Guttag, and Bruce Fischl. His research focuses on developing geometry processing and machine learning methods to advance scientific and clinical understanding in biomedical research. He obtained his PhD in Electrical Engineering and Computer Sciences from MIT, where he was advised by Polina Golland and Justin Solomon. His work has been published in top venues across computer graphics and geometry processing, machine learning, and medical image analysis. He has received the NSF GRFP and the NSERC PGS D fellowships, and is an awardee of the Siebel Fellowship and the MathWorks Fellowship.

Zoom link: https://wse.zoom.us/j/92215066845