Synthetic Data Generation for Automatic Segmentation of X-ray Computed Tomography Reconstructions of Complex...

Area of Science:

• Materials Science
• Computer Science
• Image Analysis

Background:

• Deep convolutional neural networks (DCNNs) require extensive data for microstructural segmentation of X-ray computed tomography (XCT) data.
• Manual annotation of 3D XCT data for DCNN training is prohibitively time-consuming and labor-intensive.

Purpose of the Study:

• To develop a method for generating synthetic XCT data to train DCNNs for microstructural segmentation.
• To improve the accuracy and generalizability of DCNNs trained on synthetic data for complex materials.

Main Methods:

• Generation of synthetic XCT data for a six-phase Al-Si alloy composite.
• Application of data augmentation techniques (brightness, contrast, noise, blur).
• Development and implementation of a novel deep convolutional neural network (Triple UNet) with a multi-view forwarding strategy.

Main Results:

• Achieved an overall Dice score of 0.77 for microstructural segmentation using synthetic data.
• Demonstrated the negative impact of artifacts in XCT data on segmentation accuracy when training with synthetic data.
• Validated the applicability of the proposed methods to other materials and imaging techniques.

Conclusions:

• Synthetic data generation combined with advanced DCNN architectures and augmentation strategies enables effective microstructural segmentation.
• The developed method significantly reduces the reliance on manual annotation, accelerating scientific discovery in materials science.
• Addressing data artifacts is crucial for optimizing DCNN performance in XCT-based microstructural analysis.