Abstract
High quality segmentation of diffusion tensor images (DTI) is of key interest in biomedical research and clinical application. In previous studies, most efforts have been made to construct predefined metrics for different DTI segmentation tasks. These methods require adequate prior knowledge and tuning parameters. To overcome these disadvantages, we proposed to automatically learn an adaptive distance metric by a graph based semi-supervised learning model for DTI segmentation. An original discriminative distance vector was first formulated by combining both geometry and orientation distances derived from diffusion tensors. The kernel metric over the original distance and labels of all voxels were then simultaneously optimized in a graph based semi-supervised learning approach. Finally, the optimization task was efficiently solved with an iterative gradient descent method to achieve the optimal solution. With our approach, an adaptive distance metric could be available for each specific segmentation task. Experiments on synthetic and real brain DTI datasets were performed to demonstrate the effectiveness and robustness of the proposed distance metric learning approach. The performance of our approach was compared with three classical metrics in the graph based semi-supervised learning framework.
Introduction
Diffusion Tensor Imaging (DTI) is a vital MRI technique that maps the microscopic structure of tissues, particularly white matter tracts, by measuring water molecule diffusion. Each voxel contains a tensor—a 3x3 matrix representing a diffusion ellipsoid. Segmenting these tensors is challenging because traditional distance metrics (e.g., Euclidean or Geodesic) use fixed weights for geometric shape and orientation. Since different brain structures require different weighting of these features, static metrics often fail or require tedious manual tuning.
Methods: Adaptive Graph-Based Learning
We propose a Graph-Based Semi-Supervised Learning (GSSL) model that automatically learns an adaptive Mahalanobis distance metric.
- Feature Extraction: We combine geometric features (Mean Diffusivity, Fractional Anisotropy, Volume Ratio) with orientation data (the principal eigenvector) into a discriminative distance vector.
- Adaptive Optimization: Unlike standard models with fixed similarity weights, our approach treats the weight matrix as a variable. We simultaneously optimize voxel labels and the distance metric using an iterative gradient descent method. This allows the metric to "learn" from a small set of user-provided seed labels which features are most relevant for a specific structure.
Experimental Results
The method was validated using synthetic lattices and real human brain DTI data.
Synthetic Lattice Performance
On a synthetic 15x15 lattice designed to test orientation-based separation, the adaptive approach achieved a perfect Dice Similarity Coefficient (DSC) of 1.0.
Robustness to Noise and Comparative Analysis
A key strength of the adaptive metric is its performance in noisy environments. In tests with an SNR of 10, traditional metrics like Euclidean and Geodesic distances suffered from "leaking," where the segmentation boundary bled into surrounding tissues with similar geometry but different orientation.
In real-world tests segmenting the human corpus callosum, our method maintained a high DSC of ~0.90, significantly outperforming the Geodesic (~0.80) and Euclidean (~0.70) metrics.
Conclusion
By automatically learning the optimal distance metric for specific anatomical structures, our approach provides a robust and accurate solution for DTI segmentation. While computationally more intensive than fixed-metric methods, the gain in precision and the reduction in manual parameter tuning represent a significant step forward for clinical DTI analysis. Future work will focus on GPU acceleration to reduce processing times.
Summary updated