|About this Abstract
||2020 TMS Annual Meeting & Exhibition
||Characterization of Minerals, Metals and Materials
||Using Convolutional Neural Networks to Visualize Large Serial Sectioning Datasets
||Zach Thompson, Tiberiu Stan, Peter Voorhees
|On-Site Speaker (Planned)
Serial sectioning coupled with optical microscopy can be used to obtain 3-dimensional image stacks from a materials sample quickly and with good section depth precision. The sectioning process can generate a large amount of data for a single sample (~10 GB) that needs to be accurately segmented into the phase of interest and the background. However, for certain measurements such as the number of fragments, the resulting data often contains too many artifacts that are very difficult to remove via filtering and which persist even when using state-of-the-art segmentation software. In order to remedy this, a novel machine learning approach tailored to use fewer ground truths was used to improve segmentation accuracy. Metrics are used to quantify this accuracy and what parameters most effect it. Finally, the degree of transferability of neural networks trained on one dataset and applied to other materials datasets is analyzed.