PSI - Issue 74

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

111

6

Bailly (Bailly et al., 2022) further emphasized that larger datasets do not always lead to higher accuracy but contribute to greater model stability; feature engineering and data quality remain critical factors. Nevertheless, while satisfactory performance can be achieved with very limited training data, generalization to new images, especially those acquired with different manufacturing technologies, etching conditions, or microscopy settings may be limited. Ultimately, the required dataset size depends on the specific application, as illustrated by Mikmeková (Mikmeková et al., 2023) where 230 training images were used even with pretraining on ImageNet.

Fig. 4 Comparison of segmentation accuracy and error distribution (Dice: 95.8%, IoU: 91.9%) for dislocation cells in LPBF-produced 316L steel by EOS (left) and Renishaw (right). (a) EOS original cropped image, (b) Renishaw original cropped image, (c), (d) 5 images ImageNet, (e), (f) 10 images No pretrain, (g), (h) 10 images ImageNet. (True positive: green; False positive: blue; False negative: red), Recent studies have tackled the annotation bottleneck in microstructure segmentation using either weak supervision with human-in-the-loop learning (Na et al., 2023) or synthetic dataset generation based on kinetic models (Chaurasia et al., 2023). Each approach has certain advantages: weak supervision can minimize the need for pixel-level labeling, while synthetic data enables the generation of large datasets from limited real examples. Table 3 compares the cell diameter measurements obtained by the traditional line-intercept method, the adaptive thresholding, and the U-Net segmentation. Each value in the Table 3 is based on measurements from approximately 500 individual cells per manufacturing technique and method. The differences between measuring methods are