Detection of Objects in Cryo-Electron Micrographs Using Geometric Deep Learning

Explore machine learning methods for object detection, semantic segmentation, and instance segmentation in cryo-electron micrographs. Delve into unsupervised object detection using geometric deep learning and variational autoencoders for automatic particle detection and classification in cryoEM. Discover new techniques for semantic segmentation of filaments and membranes in micrographs and tomograms, as well as a graph-based transformer for instance segmentation of point clouds. Learn how these approaches leverage geometric deep learning to build known invariants into model architectures, enhancing accuracy and data efficiency. Gain insights into the challenges of 3D instance segmentation for complex microtubule networks and the development of end-to-end tomogram analysis tools combining semantic and instance segmentation.

Intro
Object detection, semantic segmentation, and instance segmentation
Single particle cryoEM, grossly oversimplified
In natural images, objects often have unknown orientations
Spatial decoders are image generative models that are equivariant to any coordinate transformation
Prior work: spatial-VAE combined the spatial decoder with an approximate inference network to learn disentangled object representations
Spatial-VAE fails to predict uniformly distributed rotations
Convolutional neural networks are translation equivariant but not rotation equivariant
Experiment setup and evaluation
TARGET-VAE learns invariant object representations, improving semantic clustering
Dimensionless Instance Segmentation Transformer (DIST)
3D instance segmentation of complex MT networks is challenging
Combining DIST with an upstream semantic segmentation network enables end-to-end tomogram analysis (TARDIS)
Using TARDIS for fully automated semantic and instance segmentation of microtubules in situ