|About this Abstract
||MS&T21: Materials Science & Technology
||AI for Big Data Problems in Advanced Imaging, Materials Modeling and Automated Synthesis
||Optimizing the Training of Convolutional Neural Networks for Image Segmentation
||Benjamin Provencher, Aly Badran, Jonathan Kroll, Mike Marsh
|On-Site Speaker (Planned)
Recent advances in artificial intelligence (AI) have fully automated the operation of image segmentation in scientific images of many important materials samples, which was often laborious and painstaking by previous methods. But optimizing AI usage has been elusive because of the numerous parameters associated with training, in particular the question of how much training data is required. We look at training parameters for convolutional neural networks designed to segment x-ray microCT images for three different composite samples with 2, 3, and 5 material phases, respectively. Analysis of DICE performance of models trained in replicate shows the segmentation accuracy varies with the volume of provided ground-truth training data, but that the response quickly falls off. For all three samples, DICE scores of 0.95 are achieved when the volume of training data exceeds 4 megapixels. Importantly, synthetically augmented training data greatly compensates for a shortage of available natural ground-truth training data.