Abstract
Understanding the 3D microstructure of materials is essential for predicting mechanical performance. However, the standard serial sectioning method often results in 2D images plagued by physical defects and imaging artifacts. We propose WPU-Net (Weighted Propagation U-Net), a specialized convolutional neural network designed for robust boundary detection in polycrystalline materials. By incorporating a propagation module that leverages inter-slice spatial consistency and a novel Adaptive Boundary Weighted (ABW) loss, WPU-Net effectively handles noise and class imbalance. Experimental results on a new pure iron grain dataset show a 7% reduction in boundary detection error compared to existing state-of-the-art methods.
Introduction
In materials science, the quantitative analysis of microstructures relies on the "serial section" method: a sample is iteratively polished and imaged to create a stack of 2D slices for 3D reconstruction. A critical step in this pipeline is segmentation—identifying the boundaries of individual grains. This task is complicated by several factors:
- Physical Defects: Polishing often introduces scratches that mimic grain boundaries.
- Imaging Artifacts: Etching can result in blurred or missing boundaries.
- Class Imbalance: Boundary pixels are significantly outnumbered by grain interior pixels.
- Labeling Noise: Manual annotations may have slight offsets, making pixel-perfect accuracy less important than topological correctness.
Methodology
WPU-Net builds upon the U-Net architecture with two primary innovations designed to address the unique challenges of serial section imaging.
Propagation Module
Unlike standard 2D networks that process slices in isolation, WPU-Net utilizes information from the preceding slice. Since grain structures evolve gradually across the depth of a material, the segmentation of slice $i-1$ serves as a powerful prior for slice $i$. The network employs a multi-level fusion strategy, injecting this "hint" into various layers of the encoder to capture both low-level geometric features and high-level topological structures.
Adaptive Boundary Weighted (ABW) Loss
To address class imbalance and the importance of small grains, the authors introduced the ABW Loss. This loss function assigns a weight to each pixel based on its distance to the nearest boundary within its specific grain.
- Gaussian Decay: The weighting follows a Gaussian-like decay, ensuring the network focuses on the vicinity of boundaries.
- Grain-Size Adaptivity: The decay rate is faster for smaller grains, preventing them from being merged into larger neighbors during training.
- Topological Focus: By weighting the area around boundaries, the loss function encourages the model to learn the underlying logic of the grain structure, providing some tolerance for minor manual labeling errors.
Dataset: The Pure Iron Grain Dataset
The researchers introduced a benchmark dataset consisting of 296 microscopic slices of pure iron at a resolution of 1024 × 1024 pixels. This dataset includes ground truth boundaries and 3D grain labels validated by professional researchers, providing a high-quality resource for training and evaluation.
Results and Discussion
WPU-Net was evaluated against seven state-of-the-art models, including 3D U-Net and Attention U-Net.
- Quantitative Performance: WPU-Net achieved a Variation of Information (VI) score of 0.1874, representing a 7% improvement over the second-best method (3D U-Net). It also outperformed competitors in mean Average Precision (mAP) and Adjusted Rand Index (ARI).
- Qualitative Robustness: Visual analysis confirms that the propagation module allows the network to "see through" artifacts. If a boundary is visible in adjacent slices but obscured by a scratch in the target slice, the network can successfully infer its position.
- Denoising via Spatial Consistency: The multi-scale fusion of previous slice information acts as a natural denoising mechanism, significantly reducing over-segmentation caused by physical surface defects.
Conclusion
WPU-Net represents a significant advancement in automated microstructure analysis. By treating serial sections as a sequence rather than independent images, it leverages spatial redundancy to overcome the inherent noise of the imaging process. While currently focused on single-phase materials like pure iron, future work will explore multi-phase alloys and semi-supervised learning techniques to reduce the reliance on extensive manual labeling.
Summary updated