Microstructural characterization is essential in materials property control and optimization. Grains or grain boundaries are traditionally segmented by image analysis on chemically etched surfaces. Etching can be avoided by employing Backscattered Electron (BSE) imaging of Scanning Electron Microscopy (SEM). The drawback is that automated image analysis is much more complex because the technique relies only on chemical and crystallographic contrast, imposing difficulties to certain boundary detection. Deep learning-based methods represent new potentials. In this work, three Convolutional Neural Networks (CNNs), a UNet model (B), its deeply supervised variant (DS) and a Holistically-Nested Edge Detection (HED) model, were designed and assessed to detect grain boundaries using SEM-BSE images of uranium dioxide samples. In particular, a topology-preserving loss function, the clDice, was employed to enforce the predicted boundary morphology. Furthermore, a topological metric, the Betti matching error, a more appropriate metric than the traditional binary classification metrics, was used to evaluate the model performance. The results show that the DS model outperforms the others. To assess the efficiency of the DS model, we compared it with a state-of-the-art image-processing based grain boundary detection method (SEraMic) which detects boundaries via multiple SEM-BSE images of a region of interest at various tilts. Our DS model significantly improves the Betti matching error for any number of images. In particular, the DS model shows accurate results with a single input image and is even able to detect boundaries where intensity difference is below noise level. • Three CNNs were tested for grain boundary detection in SEM-BSE images of UO 2 . • A clDice loss was chosen to improve boundary morphology and topology. • Betti matching error was introduced to assess topological accuracy of predictions. • Deeply supervised UNet outperformed UNet, HED, and SEraMic method.
Zhao et al. (Mon,) studied this question.