Machine learning segmentation in 3D electron microscopy

Authors

• S Fulton
• C Baenen
• J Yang
• J Kim
• RD Leapman
• MA Aronova

Abstract

3D Electron Microscopy (EM) can be used to obtain structural information from cells and tissue at nanoscale. Application of Machine Learning (ML) can be valuable in characterizing tens of thousands of cells and their components via automated segmentation or outlining of boundaries. One challenge in automated segmentation of EM images is the determination of boundaries due to the high densities of cells or organelles. We utilized Convolutional Neural Network (CNN) with a ResNet18 backbone for automated segmentation. We employed augmentation techniques by adding noise to our labeled data, which drastically improves the robustness of our model. Another challenge in automated segmentation of densely packed cells is the lack of context when viewed in separate 2D ortho slices. We developed a 3D CNN to improve the accuracy of the segmentation, in which the context is inherent to looking at consecutive ortho slices, which enables us to surface render the segmentation results in 3D, gaining structural and spatial insight into the system we are analyzing.
We used platelets during thrombus formation as model data. Platelet differentiation and activation are fundamental to sealing leaks at sites of vascular injury during wound healing. The formation of the thrombus following a puncture wound is integral to vascular healing, however, the mechanism(s) underlying blood platelet recruitment and their activation is still poorly understood. By analyzing different regions of interest using ML, we are hoping to relate the local activation state of platelet to the global structure of the thrombus, which we have already studied.

Scientific Focus Area: Structural Biology

This page was last updated on Tuesday, August 6, 2024