A Density and Reliability Guided Aggregation for the Assessment of Vessels and Nerve Fibres Tortuosity

Pan Su, Tianhua Chen, Jianyang Xie, Baikai Ma, Hong Qi, Jiang Liu, Yitian ZhaoIEEE Access 2020 · DOI

Introduction

Tortuosity, the measure of "twistiness" in curvilinear structures, serves as a critical biomarker for various ocular and systemic conditions. In the retina, vessel tortuosity can indicate diabetic retinopathy or retinopathy of prematurity. In the cornea, nerve fiber tortuosity is linked to keratitis and peripheral neuropathy. Despite its clinical importance, the assessment of tortuosity remains largely subjective, relying on manual grading by experts. This process is not only time-consuming but also suffers from significant inter-observer variability.

While numerous mathematical metrics exist to quantify tortuosity—ranging from simple arc-chord ratios to complex curvature analysis—no single "gold standard" metric performs universally well across different image types and segmentation qualities. This research addresses this gap by proposing an automated pipeline that intelligently aggregates multiple diverse metrics into a robust, human-interpretable assessment.

The Proposed Pipeline

The core of the research is a multi-stage automated system designed to move from raw image data to clinical grades and rankings.

Overview of the proposed automated pipeline for tortuosity assessment.

The pipeline consists of four main phases: feature extraction, density-based clustering, fuzzification, and a novel aggregation scheme known as kNNDOWA. By modularizing the process, the system can handle various types of input (e.g., retinal vessels vs. corneal nerves) and provide transparent reasoning for its final outputs.

Quantifying Tortuosity: Feature Extraction

The first step involves calculating eight distinct tortuosity measures. These are categorized into three groups:

  1. Length-based metrics: Such as the arc-chord ratio.
  2. Curvature-based metrics: Analyzing the rate of change in direction.
  3. Angle-based metrics: Using slope chain codes to measure directional shifts.
Geometric representation of angle-based tortuosity metrics.

Because different metrics excel in different scenarios—for example, some are better at detecting subtle waves while others capture sharp turns—the system extracts a broad feature set to ensure no critical morphological information is lost.

Intelligent Aggregation and Fuzzification

To overcome the limitations of individual metrics, the researchers implemented a sophisticated aggregation strategy:

  • Density Peak Clustering: Instead of standard normalization, the system uses density-based clustering to identify natural groupings in the data. This approach is more resilient to outliers than traditional k-means clustering.
  • Labelled Fuzzy Sets: The extracted values are mapped to linguistic terms like "Low," "Medium," and "High." This "fuzzification" mimics clinical reasoning, where boundaries between grades are often not "crisp" but overlapping. It also makes the model an "open box," allowing clinicians to see which specific features contributed to a high tortuosity grade.
  • kNNDOWA Aggregation: The final score is calculated using k-Nearest-Neighbour guided Dependent Ordered Weighted Averaging. This data-driven mechanism assigns weights based on reliability. If one metric produces a value that significantly disagrees with the majority of other metrics for a given image, it is automatically down-weighted.

Experimental Evaluation

The system was validated across three diverse datasets: the public NERVE TORTUOSITY (NT) dataset, the RET-TORT retinal benchmark, and a large, imbalanced in-house corneal nerve dataset.

The results demonstrated that the proposed aggregation method consistently outperformed individual metrics. In four out of five test scenarios, the system achieved higher Spearman correlation coefficients than the best single measure. On the NT dataset, it reached a weighted accuracy of 0.8667, significantly surpassing conventional averaging techniques.

Conclusion

By combining density-based clustering with fuzzy logic and a reliability-guided weighting scheme, this research provides a robust framework for automated tortuosity assessment. The system not only improves accuracy over individual mathematical formulas but also offers the interpretability required for clinical decision support. Future work will focus on addressing data imbalance in clinical datasets and integrating expert feedback directly into the learned fuzzy sets.

Summary updated

BibTeX
@article{su2020density,
  author = {Pan Su and Tianhua Chen and Jianyang Xie and Baikai Ma and Hong Qi and Jiang Liu and Yitian Zhao},
  title = {{A Density and Reliability Guided Aggregation for the Assessment of Vessels and Nerve Fibres Tortuosity}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3004422},
}
Figure