Software Tools for Handling Magnetically Collected Ultra-thin Sections for Microscopy

This report presents the software tools that will be available online to help users accurately segment ultra-thin sections of brain tissue in a large image to determine the section’s coordinates. In order to predict these coordinates, the goal was to use a state-of-art object instance segmentation framework called Masked Region-based Convolutional Neural Network (Masked R-CNN) on the dataset containing sections of brain tissue in the light microscopy images. The predicted coordinates of the sections will later be used for automated image acquisition in the high-resolution electron microscope. We will demonstrate that Masked R-CNN can be used to perform highly effective and efficient automatic segmentation of microscopy images containing the sections. The machine learning pipeline used will be explained in detail. In addition, we will show the results of how different parameters and settings of this pipeline were changing the performance. This semester's project was done in collaboration with the Center for Interdisciplinary Electron Microscopy (CIME) lab and Machine Learning and Optimization (MLO) lab at EPFL.