Optimizing convolutional neural networks to perform semantic segmentation on large materials imaging datasets: X-ray tomography and serial sectioning

Machine learning was used to segment large materials science datasets resulting from synchrotron-based x-ray computed tomography (XCT) images of dendrite growth, and serial sectioning (SS) images of dendrite coarsening. Both neural networks (NNs) yielded quantitatively more accurate outputs than conventional segmentation techniques using only 30 XCT or 6 SS training images. We show that performance can be improved if NNs are trained using a large number of small images that are sampled from the fixed amount of training data. The optimal image size and number of training images was identified for the XCT and SS datasets. NN transferability was also tested by applying the highest performing XCT and SS NNs to related datasets. While the initial segmentations were successful, applying simple transformations to the raw images further improved NN performance. These results show the great predictive ability and promising future of using machine learning for segmentation of large materials science datasets.