[Work in progress] Scalable, out-of-the box segmentation of individual particles from mineral samples acquired with micro CT

Karol Gotkowski, Shuvam Gupta, Jose R. A. Godinho, Camila G. S. Tochtrop, Klaus H. Maier-Hein, Fabian Isensee

Abstract

In the fields of mineral processing and recycling, precise 3D characterization of individual particles is critical for process optimization. However, segmenting these particles from high-resolution Micro-Computed Tomography (micro-CT) volumes remains a significant challenge due to material diversity and particle proximity. This paper discusses ParticleSeg3D, an open-source deep learning framework that provides robust, "out-of-the-box" instance segmentation. By integrating the nnU-Net architecture with a specialized border-core representation and size normalization, ParticleSeg3D eliminates the need for manual parameter tuning or material-specific retraining. Evaluation across 19 diverse datasets demonstrates high generalization, achieving an F1 score of 94.84% on in-distribution samples and 93.27% on out-of-distribution materials.

Introduction

The analysis of mineral ores and recycled materials often relies on Micro-Computed Tomography (micro-CT) to provide non-destructive 3D imaging. While these scans capture intricate details, the subsequent task of instance segmentation—identifying and separating every individual particle—is fraught with difficulty. Traditional intensity-based thresholding fails when particles have overlapping density profiles or are physically touching. Standard deep learning semantic segmentation models often merge adjacent particles into single blobs. Consequently, researchers are frequently forced into labor-intensive cycles of manual annotation and model retraining for each new sample type.

Micro-CT sample and reference segmentation showing the complexity of the instance segmentation task.

Methodology

ParticleSeg3D introduces an automated pipeline designed for high scalability and generalization. The framework is built upon the nnU-Net backbone, which automatically adapts its architecture to the specific properties of the input data. To convert semantic labels into distinct instances, the model employs a "border-core" representation: it is trained to predict a central core and a surrounding border for each particle. This approach ensures that a gap is maintained between neighboring particles, facilitating separation during post-processing.

A critical component of the pipeline is the normalization of both intensity and particle size. Since micro-CT scans vary in resolution and physical particle sizes differ across samples, resizing all training patches to a common average particle size significantly enhances model robustness. Furthermore, the framework utilizes a "chunk-patch" inference strategy to process massive 3D volumes (exceeding 100GB) on standard hardware by handling overlapping sections.

The ParticleSeg3D preprocessing pipeline, including intensity and particle size normalization.

Experimental Results

The model was validated on a diverse dataset comprising natural ores (e.g., Sn-skarn, Chromite), recyclables (Lithium-ion batteries), and synthetic samples. ParticleSeg3D achieved an average F1 score of 94.84% for in-distribution samples. Notably, it maintained a high performance of 93.27% on out-of-distribution samples—materials the model had never encountered during training, such as Parisite and Zinc-rich slag.

In terms of separation quality, ParticleSeg3D drastically outperformed traditional watershed-based methods. In comparative tests, the model produced only 11 "splitters" (single particles incorrectly divided) compared to 329 produced by the baseline "ThreshWater" method. This robustness extends to imaging artifacts; the model successfully delineated boundaries even in the presence of heavy streak noise caused by high-density materials.

Discussion and Conclusion

The success of ParticleSeg3D suggests that the geometric features of "particleness" are sufficiently consistent across materials to allow for a universal representation. By prioritizing size normalization over simple intensity thresholding, the framework achieves unprecedented generalization.

However, limitations remain. The model struggles with particles smaller than 3 voxels in diameter or those thinner than the predicted border thickness. Additionally, processing large 3D volumes is computationally intensive, with median processing times around 4.4 hours. Future work will focus on optimizing inference speed and improving detection for ultra-fine particles. ParticleSeg3D is available as an open-source tool to facilitate automated, high-throughput 3D particle characterization in the mineral and recycling industries.

Summary updated

BibTeX
@inproceedings{gotkowski2023work,
  author = {Karol Gotkowski and Shuvam Gupta and Jose R. A. Godinho and Camila G. S. Tochtrop and Klaus H. Maier-Hein and Fabian Isensee},
  title = {{[Work in progress] Scalable, out-of-the box segmentation of individual particles from mineral samples acquired with micro CT}},
  booktitle = {Powder},
  year = {2023},
  doi = {10.1016/j.powtec.2023.119286},
}
Figure