Related Experiment Video
Updated: Oct 4, 2025
16:23
Fiber Connections of the Supplementary Motor Area Revisited: Methodology of Fiber Dissection, DTI, and Three Dimensional Documentation
Published on: May 23, 2017
11.4K
A reusable neural network pipeline for unidirectional fiber segmentation.
Alexandre Fioravante de Siqueira1,2, Daniela M Ushizima3,4,5, Stéfan J van der Walt3
1Berkeley Institute for Data Science, University of California, Berkeley, 94720, USA. alex.desiqueira@igdore.org.
Scientific Data
|February 3, 2022
Summary
We developed an automated computational pipeline for detecting and separating fibers in advanced ceramic-matrix composites using X-ray imaging. This open-source tool matches and sometimes surpasses human-supervised methods for analyzing these high-temperature materials.
Area of Science:
- Materials Science
- Computational Imaging
- Aerospace Engineering
Background:
- Fiber-reinforced ceramic-matrix composites are crucial for high-temperature applications, particularly in aerospace.
- Analyzing these materials requires accurate detection and separation of embedded fibers from imaging data.
- Current analysis predominantly relies on semi-supervised techniques, which can be labor-intensive and may miss subtle features.
Purpose of the Study:
- To present an open, automated computational pipeline for fiber detection in X-ray tomographic data.
- To evaluate the performance of convolutional neural network architectures for fiber separation.
- To compare the automated pipeline's efficacy against traditional semi-supervised methods.
Main Methods:
- Development of an automated computational pipeline for fiber detection from X-ray volumes.
- Application of the pipeline to a complex, non-trivial dataset.
- Testing of four distinct convolutional neural network architectures for fiber separation.
- Comparative analysis against semi-supervised segmentation techniques.
Main Results:
- The automated pipeline successfully detected and separated fibers from tomographic X-ray data.
- Convolutional neural network approaches achieved high performance metrics, with Dice and Matthews coefficients up to 98%.
- The automated methods demonstrated comparable or superior performance to semi-supervised techniques, identifying fibers missed by human-curated algorithms.
Conclusions:
- Automated computational pipelines, particularly those utilizing convolutional neural networks, offer a powerful alternative to semi-supervised methods for analyzing fiber-reinforced composites.
- The developed open-source software provides a versatile tool for researchers in materials science and related fields.
- This approach enhances the accuracy and efficiency of fiber detection and separation in complex material structures.
