[Paper Review] Instance Segmentation of Fibers from Low Resolution CT Scans via 3D Deep Embedding Learning

Tomasz Konopczyński, Thorben Kröger+2 more|arXiv (Cornell University)|Jan 4, 2019

Medical Imaging Techniques and Applications22 citations

TL;DR

This paper proposes a 3D deep embedding learning framework for instance segmentation of fibers in low-resolution CT scans of short glass fiber-reinforced polymers. By combining a 3D fully convolutional network with an embedding branch and a merging post-processing strategy, the method achieves state-of-the-art instance segmentation performance, with a merged Adjusted Rand Index of 0.6529, outperforming traditional watershed and connected components methods on challenging low-resolution data.

ABSTRACT

We propose a novel approach for automatic extraction (instance segmentation) of fibers from low resolution 3D X-ray computed tomography scans of short glass fiber reinforced polymers. We have designed a 3D instance segmentation architecture built upon a deep fully convolutional network for semantic segmentation with an extra output for embedding learning. We show that the embedding learning is capable of learning a mapping of voxels to an embedded space in which a standard clustering algorithm can be used to distinguish between different instances of an object in a volume. In addition, we discuss a merging post-processing method which makes it possible to process volumes of any size. The proposed 3D instance segmentation network together with our merging algorithm is the first known to authors knowledge procedure that produces results good enough, that they can be used for further analysis of low resolution fiber composites CT scans.

Motivation & Objective

• To address the challenge of instance segmentation in low-resolution 3D CT scans of short glass fiber-reinforced polymers, where conventional methods fail due to poor resolution.
• To develop a deep learning-based method that enables accurate extraction of individual fibers for downstream mechanical property analysis.
• To overcome limitations of Hessian-based filters and watershed segmentation in densely packed, thin fiber structures.
• To introduce a novel 3D instance segmentation framework using embedding learning that generalizes across volumetric data beyond fiber composites.
• To validate the method's robustness and scalability through a merging post-processing strategy for arbitrarily large volumes.

Proposed method

• The method employs a 3D fully convolutional network with two parallel branches: one for semantic segmentation and another for learning voxel embeddings in a continuous space.
• The embedding branch maps each voxel to a feature vector such that voxels belonging to the same fiber cluster are close in the embedding space.
• A clustering algorithm (k-means) is applied on the embedding representation within segmented regions to assign each voxel to a distinct fiber instance.
• The network processes overlapping 32×32×32 sub-volumes of the input volume to handle large-scale data, with a custom merging algorithm to resolve spatial inconsistencies between adjacent sub-volumes.
• The merging algorithm favors merging over splitting to reduce false splits, improving consistency in overlapping regions.
• The model is trained using a combined loss function combining binary cross-entropy for semantic segmentation and triplet loss for embedding learning, with hyperparameters β=1 and γ=0.001.

Experimental results

Research questions

• RQ1 Can deep embedding learning be effectively adapted to 3D instance segmentation of thin, cylindrical objects like fibers in low-resolution CT scans?
• RQ2 How does the performance of the proposed embedding-based method compare to traditional watershed and connected components techniques on low-resolution SFRP data?
• RQ3 To what extent does the quality of the semantic segmentation branch impact the final instance segmentation accuracy?
• RQ4 Can a merging strategy effectively handle overlapping sub-volume predictions without introducing significant artifacts or errors?
• RQ5 Is the proposed method scalable to arbitrarily large 3D volumes through sub-volume processing and merging?

Key findings

• The proposed method achieves a mean Adjusted Rand Index (ARI) of 0.9048 on overlapping sub-volumes, demonstrating strong per-patch performance.
• After merging overlapping sub-volumes, the ARI score decreases to 0.6529, reflecting the challenge of resolving ambiguous boundaries between closely spaced fibers.
• When using ground truth semantic segmentation instead of predicted masks, the merged ARI improves to 0.7817, indicating that the semantic segmentation branch significantly influences final instance accuracy.
• The standard connected components method performs poorly, achieving only 0.2112 merged ARI, even with ground truth semantic masks.
• The semantic segmentation branch alone achieves a Dice score of 0.9784, confirming high-quality foreground prediction.
• The method is the first known approach to produce reliable instance segmentation results from low-resolution SFRP CT scans, outperforming all baseline methods in both quantitative and qualitative evaluations.