U-SLADS: Unsupervised Learning Approach for Dynamic Dendrite Sampling

Yan Zhang, Xiang Huang, Nicola Ferrier, Emine Begum Gulsoy, Charudatta Phatak

Abstract

Novel data acquisition schemes have been an emerging need for scanning microscopy based imaging techniques to reduce the time in data acquisition and to minimize probing radiation in sample exposure. Varies sparse sampling schemes have been studied and are ideally suited for such applications where the images can be reconstructed from a sparse set of measurements. Dynamic sparse sampling methods, particularly supervised learning based iterative sampling algorithms, have shown promising results for sampling pixel locations on the edges or boundaries during imaging. However, dynamic sampling for imaging skeleton-like objects such as metal dendrites remains difficult. Here, we address a new unsupervised learning approach using Hierarchical Gaussian Mixture Models (HGMM) to dynamically sample metal dendrites. This technique is very useful if the users are interested in fast imaging the primary and secondary arms of metal dendrites in solidification process in materials science.

Introduction

High-resolution scanning microscopy, utilizing electron or X-ray beams, is a critical tool in materials science but faces a significant bottleneck: the trade-off between resolution and acquisition time. Scanning a single sample can take hours or even days, and the high-energy beams required can cause radiation damage to the specimen.

To mitigate these issues, researchers employ "sparse sampling," where only a fraction of pixels are measured and the full image is reconstructed. While existing dynamic sampling methods like SLADS (Supervised Learning Approach for Dynamic Sampling) are effective at identifying object boundaries, they often struggle with complex, branching "skeleton-like" structures, such as metal dendrites. Furthermore, supervised methods require extensive pre-labeled training datasets, which are often unavailable for novel experimental samples. U-SLADS addresses these challenges by providing a structure-aware, unsupervised framework.

Methods and Implementation

The U-SLADS algorithm treats image features as a probability distribution rather than simple intensity gradients. The core of the approach is the use of Hierarchical Gaussian Mixture Models (HGMM) to model the primary and secondary arms of a dendrite as a series of 2D Gaussian distributions.

Flowchart of the U-SLADS algorithm showing the iterative clustering and sampling process.

The process follows an iterative loop:

  1. Thresholding: Otsu’s method automatically calculates an intensity threshold to separate informative dendrite pixels from the background.
  2. Clustering: The algorithm identifies clusters of measured locations. The Bayesian Information Criterion (BIC) is used to determine the optimal number of Gaussian clusters (representing dendrite arms) at each step, preventing over-fitting.
  3. Point Selection: Using the Mahalanobis Distance, the algorithm calculates the weighted distance from unmeasured pixels to the cluster centroids, prioritizing new measurements in areas most likely to belong to the branching structure.

This hierarchical nature allows the algorithm to transition from a coarse global scan of the main "trunk" to a fine-grained local scan of secondary branches.

Experimental Results

The performance of U-SLADS was evaluated using simulated metal dendrite images derived from phase-field models. The algorithm was compared against traditional random sampling across various sampling ratios.

Quantitative comparison of PSNR and SSIM between U-SLADS and random sampling across different sampling ratios.

Quantitative Performance

As shown in the comparative plots, U-SLADS significantly outperforms random sampling in both fidelity and structural preservation:

  • Peak Signal-to-Noise Ratio (PSNR): At a 40% sampling ratio, U-SLADS achieved 11.31 dB, compared to 7.25 dB for random sampling.
  • Structural Similarity (SSIM): U-SLADS reached a score of 0.65, whereas random sampling only achieved 0.46. This indicates that U-SLADS is far more effective at preserving the intricate connectivity and branching of the dendrite skeleton.

Visual Fidelity

Qualitative analysis showed that at a mere 10% sampling ratio, U-SLADS could already identify the four primary arms of the dendrite. By the time 40% of the pixels were sampled, nearly all secondary features were accurately captured, a feat random sampling failed to replicate at the same density.

Conclusion and Future Work

U-SLADS demonstrates that unsupervised learning can effectively guide dynamic sampling for complex materials structures without the need for prior training data. This makes it a highly versatile tool for experimental physics where new, unknown samples are common.

However, the approach faces a challenge in computational overhead; as the sampling ratio increases, the time required to compute the next measurement point grows significantly. Future research will focus on optimizing these search algorithms to ensure that the "thinking time" of the software does not exceed the physical scanning speed of the microscope, potentially through more efficient binary search methods or parallelization.

Summary updated

Abstract

Novel data acquisition schemes have been an emerging need for scanning microscopy based imaging techniques to reduce the time in data acquisition and to minimize probing radiation in sample exposure. Varies sparse sampling schemes have been studied and are ideally suited for such applications where the images can be reconstructed from a sparse set of measurements. Dynamic sparse sampling methods, particularly supervised learning based iterative sampling algorithms, have shown promising results for sampling pixel locations on the edges or boundaries during imaging. However, dynamic sampling for imaging skeleton-like objects such as metal dendrites remains difficult. Here, we address a new unsupervised learning approach using Hierarchical Gaussian Mixture Mod- els (HGMM) to dynamically sample metal dendrites. This technique is very useful if the users are interested in fast imaging the primary and secondary arms of metal dendrites in solidification process in materials science.
BibTeX
@misc{zhang2018u,
  author = {Yan Zhang and Xiang Huang and Nicola Ferrier and Emine Begum Gulsoy and Charudatta Phatak},
  title = {{U-SLADS: Unsupervised Learning Approach for Dynamic Dendrite Sampling}},
  year = {2018},
  eprint = {1807.02233},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1807.02233},
}
Figure