Detection and segmentation of objects in cryo-electron micrographs using geometric deep learning

Tristan Bepler
New York Structural Biology Center (NYSBC)

In this talk I will discuss machine learning methods for object detection, semantic segmentation, and instance segmentation in cryo-electron micrographs. I will discuss recent work in unsupervised object detection using geometric deep learning and variational autoencoders to automatically detect and classify particles in cryoEM. I will then present new methods for semantic segmentation of filaments and membranes in micrographs and tomograms and a graph-based transformer for instance segmentation of point clouds that allows accurately segmenting complex, overlapping structures identified by semantic segmentation. These approaches leverage geometric deep learning techniques to explicitly build known invariants into the model architectures, improving accuracy and data efficiency of the algorithms.

Back to Workshop III: Cryo-Electron Microscopy and Beyond