A Kernel-based Approach to Diffusion Tensor and Fiber Clustering in the Human Skeletal Muscle

Radhouène Neji, Jean‐François Deux, G. Fleury, Mezri Maâtouk, Georg Langs, Jean‐Philippe Thiran, Guillaume Bassez, Alain Rahmouni, Nikolaos ParagiosSPIRE 2008

Abstract

In this report, we present a kernel-based approach to the clustering of diffusion tensors in images of the human skeletal muscle. Based on the physical intuition of tensors as a means to represent the uncertainty of the position of water protons in the tissues, we propose a Mercer (i.e. positive definite) kernel over the tensor space where both spatial and diffusion information are taken into account. This kernel highlights implicitly the connectivity along fiber tracts. We show that using this kernel in a kernel-PCA setting compounded with a landmark-Isomap embedding and k-means clustering provides a tractable framework for tensor clustering. We extend this kernel to deal with fiber tracts as input using the multi-instance kernel by considering the fiber as set of tensors centered in the sampled points of the tract. The obtained kernel reflects not only interactions between points along fiber tracts, but also the interactions between diffusion tensors. We give an interpretation of the obtained kernel as a comparison of soft fiber representations and show that it amounts to a generalization of the Gaussian kernel Correlation. As in the tensor case, we use the kernel-PCA setting and k-means for grouping of fiber tracts. This unsupervised method is further extended by way of an atlas-based registration of diffusion-free images, followed by a classification of fibers based on non-linear kernel Support Vector Machines (SVMs) and kernel diffusion. The experimental results on a dataset of diffusion tensor images of the calf muscle of 25 patients (of which 5 affected by myopathies, i.e. neuromuscular diseases) show the potential of our method in segmenting the calf in anatomically relevant regions both at the tensor and fiber level.

Introduction

Diffusion Tensor Imaging (DTI) is a powerful MRI technique that maps the diffusion of water molecules in tissues. While widely used to study white matter in the brain, applying it to skeletal muscle presents unique challenges. Muscle tissue consists of complex, ordered structures called myofibers, and understanding their organization is crucial for diagnosing and monitoring neuromuscular diseases (myopathies).

The core challenge addressed in this work is clustering: the grouping of millions of individual data points (tensors) or reconstructed pathways (fiber tracts) into distinct muscle groups like the soleus or gastrocnemius. Existing methods often rely on simple distances that ignore the directional nature of the data or fail to account for the "flow" of connectivity along a fiber. This paper introduces a mathematically rigorous similarity measure that respects the underlying physics of water diffusion.

Methodology

The Tensor Kernel

Instead of using a standard distance metric, we propose an "expected likelihood kernel." This kernel treats each tensor as a probability distribution of water proton positions. It calculates the similarity between two points not just by spatial proximity, but by the likelihood of a water molecule moving from one to the other given local diffusion properties. This makes the kernel anisotropic, naturally favoring connections along the direction of the muscle fibers.

Dimensionality Reduction and Clustering

To manage the high-dimensional tensor information, we employ Kernel PCA combined with Landmark-Isomap. This approach embeds the data into a lower-dimensional Euclidean space while preserving the "diffusive pathways" (the shortest paths through the data). This unsupervised method is further extended for practical use via an atlas-based approach using Support Vector Machines (SVMs) to learn from expert-segmented data, ensuring spatially smooth and anatomically consistent results.

Fiber Clustering

We extend the tensor kernel to entire fibers using a "summation kernel" (a multi-instance learning approach). By treating a fiber as a set of tensors, the algorithm can compare the "soft representations" of two different fiber tracts. This generalization of Gaussian kernel Correlation allows for the grouping of fiber tracts into anatomically relevant bundles.

Experimental Results

Dataset and Setup

The study utilized a clinical dataset of 25 participants: 20 healthy volunteers and 5 patients affected by myopathies. DTI of the calf muscle was performed on a 1.5T MRI scanner. High-resolution T1-weighted images were manually segmented by an expert into 7 distinct muscle groups to provide a benchmark for evaluation.

Key Findings

  • Quantitative Accuracy: The unsupervised fiber clustering achieved a mean Dice coefficient of approximately 80% across all subjects compared to manual segmentation.
  • Clinical Relevance: The method successfully segmented muscles in patients with myopathies, where fat replacement often obscures traditional imaging signals.
  • Robustness: By projecting data into a feature space via Kernel PCA, the method effectively denoises the DTI field, significantly outperforming standard spectral clustering in noisy regions near joints.
  • Implicit Connectivity: The kernel embedding reveals the complex "bipennate" (feather-like) architecture of muscles like the soleus without requiring explicit tractography for every fiber.

Conclusion

This kernel-based framework provides a sophisticated tool for "virtual dissection," allowing for the quantitative study of specific muscle groups. By bridging the gap between raw tensor data and high-level anatomy, it offers a robust means to monitor changes due to exercise, aging, or neuromuscular disease. Future work may explore Gaussian Processes to model fibers as continuous curves for even more precise soft representations of muscle architecture.

Summary updated

Abstract

In this report, we present a kernel-based approach to the clustering of diffusion tensors in images of the human skeletal muscle. Based on the physical intuition of tensors as a means to represent the uncertainty of the position of water protons in the tissues, we propose a Mercer (i.e. positive definite) kernel over the tensor space where both spatial and diffusion information are taken into account. This kernel highlights implicitly the connectivity along fiber tracts. We show that using this kernel in a kernel-PCA setting compounded with a landmark-Isomap embedding and k-means clustering provides a tractable framework for tensor clustering. We extend this kernel to deal with fiber tracts as input using the multi-instance kernel by considering the fiber as set of tensors centered in the sampled points of the tract. The obtained kernel reflects not only interactions between points along fiber tracts, but also the interactions between diffusion tensors. We give an interpretation of the obtained kernel as a comparison of soft fiber representations and show that it amounts to a generalization of the Gaussian kernel Correlation. As in the tensor case, we use the kernel-PCA setting and k-means for grouping of fiber tracts. This unsupervised method is further extended by way of an atlas-based registration of diffusion-free images, followed by a classification of fibers based on non-linear kernel Support Vector Machines (SVMs) and kernel diffusion. The experimental results on a dataset of diffusion tensor images of the calf muscle of 25 patients (of which 5 affected by myopathies, i.e. neuromuscular diseases) show the potential of our method in segmenting the calf in anatomically relevant regions both at the tensor and fiber level.
BibTeX
@inproceedings{neji2008kernel,
  author = {Radhouène Neji and Jean‐François Deux and G. Fleury and Mezri Maâtouk and Georg Langs and Jean‐Philippe Thiran and Guillaume Bassez and Alain Rahmouni and Nikolaos Paragios},
  title = {{A Kernel-based Approach to Diffusion Tensor and Fiber Clustering in the Human Skeletal Muscle}},
  booktitle = {SPIRE},
  year = {2008},
}
Figure