Automated Vessel Segmentation Using Infinite Perimeter Active Contour Model with Hybrid Region Information with Application to Retinal Images

Yitian Zhao, Lavdie Rada, Ke Chen, Simon P. Harding, Yalin ZhengIEEE TMI 2015 · DOI

Introduction

Mapping blood vessels in the retina is a critical diagnostic step for identifying systemic diseases such as diabetes, hypertension, and stroke. However, automated segmentation remains a significant challenge due to the thin, branching, and "oscillatory" nature of retinal vasculature. Traditional Active Contour Models (ACMs) typically employ a regularization term based on the Hausdorff measure (boundary length). While effective for smoothing noise, this "shortening flow" often causes the model to ignore or snap off the very thin capillaries that are clinically essential.

Methodology

This paper introduces the Infinite Perimeter Active Contour with Hybrid Region Information (IPACHI), an unsupervised model designed to preserve complex geometric structures. The model incorporates two primary innovations:

  1. Infinite Perimeter Regularizer: By replacing the standard length-based penalty with the Lebesgue measure of a $\gamma$-neighborhood around the boundary, the model can capture fractal-like and highly oscillatory structures. This allows the contour to evolve into thin branches without the smoothing bias of traditional models.
  2. Hybrid Region Information: The model integrates raw pixel intensity with a Local Phase-based enhancement map. Local phase is intensity-invariant, enabling the detection of vessel edges even under uneven illumination or the "central light reflex" (a common bright streak in retinal vessels), while raw intensity ensures the vessel body is correctly identified.
Figure 1: Visualization of the image enhancement process across different retinal image types, showing (A) original images, (B) vessel enhancement, (C) local phase information, and (D) the combined hybrid map.

Experimental Evaluation

The IPACHI model was evaluated on three diverse public benchmarks to test its robustness across different imaging modalities:

  • DRIVE: 40 color fundus images, the standard for vessel segmentation.
  • STARE: 20 color fundus images, many featuring significant pathological anomalies.
  • VAMPIRE: 8 ultra-wide-field retinal angiographic images, characterized by high resolution and a large field of view.

Results and Discussion

The IPACHI model consistently outperformed other unsupervised methods, such as Frangi filters (FR) and wavelet transforms (WL), across all datasets. The use of Local Phase (LP) enhancement was a key driver of this performance.

Figure 3: Performance metrics (Sensitivity, Specificity, Accuracy, AUC, and Dice Coefficient) for the DRIVE, STARE, and VAMPIRE datasets comparing different enhancement techniques.

On the DRIVE dataset, the model achieved a sensitivity of 0.742 and an accuracy of 0.954, nearing the performance of human experts. In the STARE dataset, it maintained a high sensitivity of 0.780 despite the presence of pathologies. Furthermore, its performance on the VAMPIRE dataset demonstrates scalability to ultra-wide-field imagery, surpassing specialized methods like the "VP" algorithm.

Conclusion

The IPACHI model demonstrates that a mathematically rigorous unsupervised approach can rival complex supervised deep learning methods without requiring large labeled training sets. By embracing the "infinite perimeter" of biological vasculature rather than forcing smoothness, the model provides a more accurate map of the retinal rivers. Future work will focus on optimizing computational efficiency for real-time clinical use and extending the mathematical framework to 3D imaging modalities like CT and MRI.

Summary updated

BibTeX
@article{zhao2015automated,
  author = {Yitian Zhao and Lavdie Rada and Ke Chen and Simon P. Harding and Yalin Zheng},
  title = {{Automated Vessel Segmentation Using Infinite Perimeter Active Contour Model with Hybrid Region Information with Application to Retinal Images}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2015},
  doi = {10.1109/tmi.2015.2409024},
}
Figure