Explore computer vision methods for automated segmentation and analysis of subcellular components in cryo-electron tomography (cryoET) data. Delve into the challenges of analyzing 3D tomogram data and learn about approaches for segmenting mitochondrial volumes and granules in iPSC-derived neurons from Huntington's Disease patient samples. Discover ongoing research on label-efficient segmentation techniques to scale computer vision analysis to larger numbers of features. Examine the potential of unsupervised learning methods, including hyperbolic space models and domain adaptive region-based convolutional neural networks, for biomedical image segmentation. Gain insights into the future directions of computer vision applications in cryoET data analysis and their implications for understanding subcellular structures and interactions.
Computer Vision for Segmenting & Analyzing Subcellular Components in Cryo-Electron Tomography
Institute for Pure & Applied Mathematics (IPAM) via YouTube
Overview
Syllabus
Intro
CryoET data analysis offers many opportunities... but also present significant challenges
Quantifying mitochondrial granules in HD neurons
HD Patient iPSC-Derived Neurons on CryoEM Grids
Objective: Automated Quantification of Mito Granules
Training a 3D Segmentation Model
Quantitative Characterization of Mitochondrial Granules in Neurites of HD Neurons
Ongoing and Future Directions
Further reducing annotation needs for training computer vision algorithms
Unsupervised learning methods for segmentation
Motivations for our work in unsupervised segmentation
Hyperbolic space
Poincare ball model
Gyrovector operations bring linear algebra to the Poincare ball
Our work: reconstructing visual hierarchy as a pretext task
Self-supervised hierarchical triplet loss
Hyperbolic clustering
Experimental validation
BraTS Dataset (Menze et al 2015)
Cryogenic electron tomography
Summary
Domain adaptive unsupervised instance segmentation for biomedical images
DARCNN: Domain Adaptive Region-based Convolutional Neural Network for Unsupervised Instance Segmentation in Biomedical Images
Background: Mask R-CNN instance segmentation architecture
DARCNN model: feature-level adaptation
DARCNN model: pseudo-labelling
Potential unsupervised discovery applications in Cryo-ET
Developing Computer Vision Methods for Segmenting and Analyzing Subcellular Components in Cryo-ET
Taught by
Institute for Pure & Applied Mathematics (IPAM)
