PSI - Issue 74

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

110

5

Hierarchies) algorithm based on two geometric parameters: cell area and the squared ratio of the major to minor axes, which enhances the separation between structural types (Fig. 3b–c). This clustering process distinguished three main categories: clustered cells (Fig. 3d), hexagonal cells (Fig. 3e), and elongated cells (Fig. 3f). Clustered cells typically arise from thresholding artifacts, often due to incomplete or irregular cell boundaries caused by uneven etching. Despite postprocessing efforts, some of these artifacts persist. While structurally distinct, elongated cells are not the focus of the final analysis. Only hexagonal cells, characterized by their regular morphology, were selected for quantitative size and density measurements. Thus, the BIRCH algorithm served as a classification tool to distinguish morphologically distinct cell groups (cluster, hexagonal, elongated), ensuring that measurements were performed exclusively on relevant structures.

Fig. 3 Image processing of 3D printed 316L microstructure (a) raw image, (b) visualization of highlighted distinguished groups in plot, (c) representation of grouped data in the original image, (d) clustered cells, (e) hexagonal-shaped cells, (f) elongated cells

3.2 Introducing U -Net In the case of the EOS-manufactured microstructure, the parameters for adaptive thresholding were carefully fine tuned to achieve precise cell segmentation. However, applying the same parameters to the Renishaw-produced structures resulted in less accurate segmentation. This discrepancy may be attributed to differences in powder composition, printing technology, and processing parameters, all of which may influence the microstructural features and the response of the material during chemical etching. In particular, some cell boundaries in the Renishaw samples were not as distinctly revealed as those in the EOS samples, leading to a reduced number of clearly segmentable cells. Fig. 4 presents the segmentation outcomes using both pretrained and non-pretrained U-Net models. Model performance (see Fig. 4b–g) was evaluated with two widely used metrics: the Dice Similarity Coefficient (DICE) and Intersection over Union (IoU), both of which measure the overlap between predicted and ground-truth masks. The comparison highlights the benefit of using pretrained encoders, especially in cases where thresholding-based preprocessing is challenged by morphological variability (Fig 4a) or suboptimal contrast (Fig. 4b). The highest segmentation accuracy was achieved for the EOS-manufactured structure (Fig. 4g) when the U-Net was trained on ten images and employed a ResNet encoder pretrained on ImageNet. Notably, pretraining (Fig 4c, d) had a more significant impact on performance than simply increasing the number of training images (Fig 4e, f). Similar conclusions were drawn by Stuckner (Stuckner et al., 2022) who showed that high IoU scores can be achieved even with a single training image, and by Sevi and Aydin (Sev ı̇ and Aydin, 2023) who highlighted the effectiveness of U Net for small datasets, contradicting the common assumption that deep learning always requires large amounts of data.