Syntéza snímků precipitátů pro segmentační neuronovou síť

• Název práce v češtině: Syntéza snímků precipitátů pro segmentační neuronovou síť
• Název v anglickém jazyce: Image Synthesis for Precipitate Segmentation Neural Network
• Klíčová slova: Segmentace precipitatů|generování syntetických obrazů|analýza mikroskopických snímků|neuronové sítě|hluboké učení|materiálová věda|segmentace obrazů|analýza mikrostruktury
• Klíčová slova anglicky: Precipitate segmentation|synthetic image generation|microscopy image analysis|neural networks|deep learning|materials science|image segmentation|microstructure analysis
• Akademický rok vypsání: 2024/2025
• Typ práce: diplomová práce
• Ústav: Katedra softwaru a výuky informatiky (32-KSVI)
• Vedoucí / školitel: RNDr. Jan Blažek, Ph.D.
• Řešitel: skrytý - zadáno a potvrzeno stud. odd.
• Datum přihlášení: 24.04.2025
• Datum zadání: 05.05.2025
• Datum potvrzení stud. oddělením: 05.05.2025
• Konzultanti: Mgr. Jaroslav Knotek

Zásady pro vypracování

High-energy irradiation of metals in power plants can lead to the formation of precipitates – alterations in crystal structure that compromise material integrity and performance. Detection and segmentation of these precipitates from microscopic images is critical for material analysis. However, due to significant visual diversity across these images, this task is too complex for classical image processing algorithms. Neural networks offer a more robust solution, though they require extensive labeled training datasets, which are costly to obtain through manual annotation.
This thesis explores expanding the training dataset through synthetic image generation to address the shortage of real labeled images. An image generation pipeline will be designed and implemented to produce realistic synthetic microscopy images that capture the statistical and structural properties of real-world samples, including precipitate morphology, grain boundaries, and noise patterns.
The impact of these synthetic images will be evaluated and the segmentation performance of models trained on fully synthetic, mixed and real data will be quantitatively assessed. Recommended evaluation metrics include:
• Dice Similarity Coefficient (DSC)
• Intersection over Union (IoU)
• Precision and Recall

The outcomes of this research will offer insights into the effectiveness of synthetic data for training deep learning models in data-scarce domains and provide practical guidance on the optimal combination of real and synthetic datasets.

A literature review will cover key topics such as:
• Precipitate segmentation and microstructure analysis in materials science
• Synthetic data generation techniques (e.g., procedural modeling, GANs, domain randomization)
• Performance of neural networks on medical or microscopy image segmentation tasks
• Benchmarking strategies and metrics for evaluating image segmentation quality

Seznam odborné literatury

DaCosta, L. R., Sytwu, K., Groschner, C. K., & Scott, M. C. (2024). A robust synthetic data generation framework for machine learning in high-resolution transmission electron microscopy (HRTEM). npj Computational Materials, 10(1). https://doi.org/10.1038/s41524-024-01336-0
Trampert, P., Rubinstein, D., Boughorbel, F., Schlinkmann, C., Luschkova, M., Slusallek, P., Dahmen, T., & Sandfeld, S. (2021). Deep neural networks for analysis of microscopy images—Synthetic data generation and adaptive sampling. Crystals, 11(3), 258. https://doi.org/10.3390/cryst11030258
Horwath, J.P., Zakharov, D.N., Mégret, R. et al. Understanding important features of deep learning models for segmentation of high-resolution transmission electron microscopy images. npj Comput Mater 6, 108 (2020). https://doi.org/10.1038/s41524-020-00363-x
Lin, B., Emami, N., Santos, D. A., Luo, Y., Banerjee, S., & Xu, B.-X. (2022). A deep learned nanowire segmentation model using synthetic data augmentation. npj Computational Materials, 8(1), 88. https://doi.org/10.1038/s41524-022-00767-x