PSI - Issue 74

Tomáš Vražina et al. / Procedia Structural Integrity 74 (2025) 106 –113 Tomáš Vražina / Structural Integrity Procedia 00 (202 5 ) 000 – 000

112

7

relatively small when the standard deviation is considered, indicating that all three approaches provide consistent results for average cell diameter. However, while the line-intercept method is limited to average diameter measurements, both the adaptive thresholding and the U-Net enable the extraction of additional information, such as cell eccentricity, precise area, distribution statistics, and cell count. It is important to note that the adaptive thresholding can be affected by microstructural artifacts, as illustrated in Fig. 3d–f, which may have an impact on the accuracy of the measurement. In contrast, the U-Net approach demonstrates improved robustness to such artifacts, enabling more reliable quantification of individual cell features. Table 3. Comparison of cell diameter (µm) obtained by various measurement methods for different manufacturing techniques. Methodology EOS RENISHAW Line-intercept 0.5 ± 0.11 0.57 ± 0.12 Adaptive thresholding 0.6 ± 0.05 0.47 ± 0.03 U-net 0.62 ± 0.03 0.52 ± 0.03 4. Conclusions Based on the comprehensive analysis of the 316L cellular microstructure using a combination of classical and deep learning-based segmentation methods, the following key conclusions can be drawn: • A hybrid approach combining adaptive thresholding with deep learning (U-Net with a pretrained ResNet34 encoder) enables reliable and efficient segmentation of hexagonal dislocation cells in LPBF-produced 316L austenitic steel, applicable to samples produced by both EOS and Renishaw machines. • All three methods under investigation provided similar results concerning cell diameter. Nevertheless, Adaptive thresholding and U-net were faster and offered additional parameters from analyzed pattern. • U-Net models utilizing ImageNet-pretrained encoders achieved higher segmentation accuracy, with a stronger impact than merely expanding the training set or training without preexisting weights. • The study confirms that, although deep learning models can perform well with small datasets in specific tasks, their generalizability to images acquired under different manufacturing or imaging conditions may require further data augmentation or adaptation. Acknowledgements The authors would like to express their gratitude for the financial support of the Czech Science Foundation within the project No. GA23-05372S. This work has also been supported by the project INTER-COST No. LUC24093 funded by the Ministry of Education, Youth and Sports of the Czech Republic. Support from the Czech Academy of Sciences within the framework of the project Lumina quaeruntur is acknowledged. References Babinský, T., Šu lák, I., Kuběna, I., Man, J., Weiser, A., Švábenská, E., Englert, L., Guth, S., 2023. Thermomechanical fatigue of additively manufactured 316L stainless steel. Materials Science and Engineering: A 869, 144831. Bailly, A., Blanc, C., Francis, É., Guillotin, T., Jamal, F., Wakim, B., Roy, P., 2022. Effects of dataset size and interactions on the prediction performance of logistic regression and deep learning models. Computer Methods and Programs in Biomedicine 213, 106504. Chaurasia, N., Jha, S.K., Sangal, S., 2023. A novel training methodology for phase segmentation of steel microstructures using a deep learning algorithm. Materialia 30, 101803. Chen, Y., Wang, X., shen, J., Peng, Y., Jiang, Y., Yang, X., Leen, S.B., Gong, J., 2022. Deformation mechanisms of selective laser melted 316L austenitic stainless steel in high temperature low cycle fatigue. Materials Science and Engineering: A 843, 143123. He, K., Zhang, X., Ren, S., Sun, J., 2015. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification, in: 2015 IEEE International Conference on Computer Vision (ICCV). Presented at the 2015 IEEE International Conference on Computer Vision (ICCV), pp. 1026–1034. Hu, Z., Gao, S., Zhang, L., Shen, X., Seet, H.L., Nai, S.M.L., Wei, J., 2022. Micro laser powder bed fusion of stainless steel 316L: Cellular structure, grain characteristics, and mechanical properties. Materials Science and Engineering: A 848, 143345. Kong, D., Dong, C., Wei, S., Ni, X., Zhang, L., Li, R., Wang, L., Man, C., Li, X., 2021. About metastable cellular structure in additively manufactured austenitic stainless steels. Additive Manufacturing 38, 101804.