Grain size measurement of calcium phosphate ceramic using deep learning

This is the authors' abstract. We don't add key points for this paper.

Grain size determination is crucial for correlating microstructure with material properties in ceramics, but manual measurement is time-consuming and operator dependent. This study presents a neural network (U-Net) for automated grain boundary segmentation in scanning electron microscopy (SEM) images of β-tricalcium phosphate ceramics. Two loss functions, binary cross-entropy (BCE) and binary focal cross-entropy (BFCE), were evaluated. High segmentation performance was achieved across 70 validation images of varying quality, with intersection-over-union (IoU) scores of 0.953 (BCE) and 0.943 (BFCE). No image enhancement methods were necessary. The automated approach achieved higher precision than manual measurements (relative standard deviations < 1.5%). Linear regression confirmed strong correlation between automated and manual measurements. Although no ground truth is available, systematic differences can be addressed through dataset-specific calibration. The workflow provides a robust and reproducible framework for automated grain size analysis, remaining effective in the presence of intra-grain topography, grain fractures, and blurred grain boundaries.