Quantification of retinal blood leakage in fundus fluorescein angiography in a retinal angiogenesis model

César H. Comin, Demetrios I. Tsirukis, Ye Sun, Xiaoyin XuScientific 2021 · DOI

TL;DR

This paper introduces a novel computational pipeline for automatically quantifying retinal blood leakage and blood vessels in fundus fluorescein angiography (FFA) images from a preclinical mouse model (Vldlr<sup>-/-</sup>). This method addresses the current lack of objective analysis for this model, enabling more consistent and efficient drug screening and mechanistic studies in vascular eye diseases.

What Problem Does This Paper Address?

Vascular eye diseases, such as age-related macular degeneration and diabetic retinopathy, are characterized by abnormal blood vessel growth and leakage in the retina. Preclinical animal models are crucial for developing new treatments, and the Vldlr<sup>-/-</sup> mouse model is particularly valuable due to its similarities to human neovascular AMD. Fundus fluorescein angiography (FFA) is a standard imaging technique to visualize this leakage. However, interpreting FFA images from the Vldlr<sup>-/-</sup> model is challenging because existing methods for quantifying leakage are often manual, subjective, and time-consuming. This lack of automated, objective quantification hinders efficient drug screening and detailed mechanistic studies.

Methods

The authors developed a three-task computational pipeline:

  1. Blood Leakage Identification: This involves separating leakage regions from blood vessels. The process starts with removing small blood vessels using grayscale opening operations with line structuring elements. Then, large blood vessels are identified and removed using a combination of morphological operations and quadrature filters. The final leakage image is obtained by subtracting the segmented blood vessels from the original image.
  2. Blood Vessel Segmentation: This task utilizes a set of log-Gabor quadrature filters, designed for optimal Fourier domain coverage, to enhance blood vessel structures. A vesselness map is generated, followed by a Chan-Vese graph cut segmentation to precisely delineate the blood vessels.
  3. Image Registration: To account for sample movement and variations across time-series FFA images, a novel optic disk detection algorithm based on graph analysis of blood vessel medial axes is employed. This algorithm identifies a stable reference point, which is then used to register images to a fixed coordinate frame, typically the first image in the sequence, using a rigid transformation and cross-correlation similarity.

Datasets

The study used fundus fluorescein angiography (FFA) images from Vldlr<sup>-/-</sup> mice. These images were acquired at different time points after fluorescein injection. For evaluation, two specialists manually annotated leakage regions in nine images to create ground truth datasets.

Key Insights and Ideas

  • Bridging the Gap in Preclinical Models: The Vldlr<sup>-/-</sup> mouse model is a valuable tool, but the absence of automated quantification methods has been a bottleneck. This work provides a much-needed objective tool.
  • Separating Leakage from Vessels is Key: The core challenge is distinguishing subtle leakage signals from the much brighter blood vessels. The pipeline's success hinges on effectively removing vessels without damaging leakage information.
  • Graph Analysis for Stable Reference Points: Developing a novel method to find a consistent reference point (near the optic disk) using blood vessel topology is crucial for accurate image registration, especially in dynamic in vivo imaging.
  • Quantitative Validation is Paramount: Comparing the automated method against manual annotations by multiple experts is essential for demonstrating its reliability and establishing its performance relative to human variability.

Key Results

  • The developed method successfully segmented blood vessels and identified leakage regions.
  • The pipeline achieved a relative difference of approximately 10% ± 6% for leakage intensity and 17% ± 8% for leakage area when compared to ground truth quantifications.
  • Pearson correlation coefficients between the method's results and ground truth were high: 0.98 for leakage intensity and 0.94 for leakage region.
  • The method demonstrated consistent measurements across different mice, showing similar quantification of retinal leakages over time.

Limitations and Future Work

  • Artifacts and Non-Rigid Registration: The authors acknowledge that artifacts in FFA images (e.g., from specific diseases or imaging issues) might be misidentified as leakage. They also note that a rigid transformation for registration might not be sufficient in all cases, suggesting that non-rigid registration could be explored.
  • Future Directions: The authors propose several avenues for future work:
  • Calculating more specific properties, such as the relationship between leakage intensity and blood vessel length, or leakage region and blood vessel diameter.
  • Integrating deep learning techniques to potentially improve performance and handle complex image characteristics.
  • Expanding the application to other preclinical models or even human FFA images, with appropriate adjustments.

Summary updated

Abstract

Blood leakage from the vessels in the eye is the hallmark of many vascular eye diseases. One of the preclinical mouse models of retinal blood leakage, the very-low-density-lipoprotein receptor deficient mouse (Vldlr<sup>-/-</sup>), is used for drug screening and mechanistic studies. Vessel leakage is usually examined using Fundus fluorescein angiography (FFA). However, interpreting FFA images of the Vldlr<sup>-/-</sup> model is challenging as no automated and objective techniques exist for this model. A pipeline has been developed for quantifying leakage intensity and area including three tasks: (i) blood leakage identification, (ii) blood vessel segmentation, and (iii) image registration. Morphological operations followed by log-Gabor quadrature filters were used to identify leakage regions. In addition, a novel optic disk detection algorithm based on graph analysis was developed for registering the images at different timepoints. Blood leakage intensity and area measured by the methodology were compared to ground truth quantifications produced by two annotators. The relative difference between the quantifications from the method and those obtained from ground truth images was around 10% ± 6% for leakage intensity and 17% ± 8% for leakage region. The Pearson correlation coefficient between the method results and the ground truth was around 0.98 for leakage intensity and 0.94 for leakage region. Therefore, we presented a computational method for quantifying retinal vascular leakage and vessels using FFA in a preclinical angiogenesis model, the Vldlr<sup>-/-</sup> model.
BibTeX
@inproceedings{comin2021quantification,
  author = {César H. Comin and Demetrios I. Tsirukis and Ye Sun and Xiaoyin Xu},
  title = {{Quantification of retinal blood leakage in fundus fluorescein angiography in a retinal angiogenesis model}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-99434-2},
}
Figure