Automated Detection of Vessel Abnormalities on Fluorescein Angiogram in Malarial Retinopathy

Yitian Zhao, Ian J. C. MacCormick, David Parry, Nicholas A. V. Beare, Simon Harding, Yalin ZhengScientific 2015 · DOI

Abstract

The detection and assessment of intravascular filling defects is important, because they may represent a process central to cerebral malaria pathogenesis: neurovascular sequestration. We have developed and validated a framework that can automatically detect intravascular filling defects in fluorescein angiogram images. It first employs a state-of-the-art segmentation approach to extract the vessels from images and then divide them into individual segments by geometrical analysis. A feature vector based on the intensity and shape of saliency maps is generated to represent the level of abnormality of each vessel segment. An AdaBoost classifier with weighted cost coefficient is trained to classify the vessel segments into normal and abnormal categories. To demonstrate its effectiveness, we apply this framework to 6,358 vessel segments in images from 10 patients with malarial retinopathy. The test sensitivity, specificity, accuracy, and area under curve (AUC) are 74.7%, 73.5%, 74.1% and 74.2% respectively when compared to the reference standard of human expert manual annotations. This performance is comparable to the agreement that we find between human observers of intravascular filling defects. Our method will be a powerful new tool for studying malarial retinopathy.

Introduction: The Challenge of Cerebral Malaria

Cerebral malaria (CM) remains a significant cause of pediatric mortality and morbidity in sub-Saharan Africa. A defining feature of its pathogenesis is "sequestration," where parasitized red blood cells adhere to the endothelium of small blood vessels, leading to microvascular obstruction. Because the retina is embryologically and anatomically an extension of the central nervous system, malarial retinopathy (MR) serves as a critical clinical proxy for observing these neurovascular processes in vivo.

A key indicator of sequestration in the retina is the presence of intravascular filling defects (IVFDs). On fluorescein angiograms (FA), these appear as irregular "mottling" or small "bites" taken out of the vessel walls, indicating disrupted blood flow.

Fluorescein angiogram (FA) images showing intravascular filling defects (IVFDs). Arrows indicate irregular vessel walls and mottling characteristic of malarial retinopathy.

Despite their clinical importance, identifying IVFDs manually is a grueling task, often requiring over an hour per image, and is subject to significant inter-observer variability. There is an urgent need for automated, objective tools to quantify these lesions for both clinical diagnosis and research.

Methodology: A Saliency-Based Detection Pipeline

The researchers developed a multi-stage automated framework to detect IVFDs:

  1. Vessel Segmentation: The vascular tree is extracted from noisy FA images using a combination of local phase enhancement and a graph-cut-based Chan-Vese model.
  2. Geometric Analysis: The extracted vasculature is thinned to its centerlines and partitioned into individual segments at every branch point.
  3. Saliency Map Generation: This core innovation treats IVFDs as "salient" regions—areas that are statistically "surprising" relative to their surroundings. The system generates intensity-based maps (using Kernel Density Estimation to find unusual brightness) and shape-based maps (measuring abrupt changes in vessel diameter or curvature).
  4. Classification: A feature vector of 21 descriptors is extracted from these maps and processed by an AdaBoost classifier. A weighted cost coefficient was employed to address the class imbalance between normal and abnormal segments.

Results and Performance

The framework was validated on a dataset of 10 retinal FA images from pediatric patients, comprising 6,358 individual vessel segments. The system achieved a sensitivity of 74.7%, a specificity of 73.5%, and an overall accuracy of 74.1%. The Area Under the Curve (AUC) was 74.2%.

Automated classification of retinal vessel segments. Red lines indicate detected abnormalities (IVFDs), while green lines represent normal vessel segments.

Notably, the system's performance was most robust in small vessels (AUC of 0.776), which are the primary sites of sequestration in cerebral malaria. Furthermore, the algorithm's performance metrics are statistically comparable to the level of agreement typically found between two human experts, suggesting that the tool provides human-level diagnostic consistency.

Discussion and Future Directions

By framing pathology as "visual surprise" (saliency), the researchers bypassed the need for rigid morphological definitions of lesions, allowing for the detection of diverse irregular textures. This reinforces the value of the "neuro-retinal" link in malaria research.

While the results are promising, the study is limited by its small patient sample size (n=10) and the inherent difficulty of imaging the smallest peri-capillary vessels with sufficient contrast. Future work will focus on expanding the dataset and exploring the application of this saliency-based framework to other neurovascular conditions, such as diabetic retinopathy.

Summary updated

Abstract

The detection and assessment of intravascular filling defects is important, because they may represent a process central to cerebral malaria pathogenesis: neurovascular sequestration. We have developed and validated a framework that can automatically detect intravascular filling defects in fluorescein angiogram images. It first employs a state-of-the-art segmentation approach to extract the vessels from images and then divide them into individual segments by geometrical analysis. A feature vector based on the intensity and shape of saliency maps is generated to represent the level of abnormality of each vessel segment. An AdaBoost classifier with weighted cost coefficient is trained to classify the vessel segments into normal and abnormal categories. To demonstrate its effectiveness, we apply this framework to 6,358 vessel segments in images from 10 patients with malarial retinopathy. The test sensitivity, specificity, accuracy, and area under curve (AUC) are 74.7%, 73.5%, 74.1% and 74.2% respectively when compared to the reference standard of human expert manual annotations. This performance is comparable to the agreement that we find between human observers of intravascular filling defects. Our method will be a powerful new tool for studying malarial retinopathy.
BibTeX
@inproceedings{zhao2015automated,
  author = {Yitian Zhao and Ian J. C. MacCormick and David Parry and Nicholas A. V. Beare and Simon Harding and Yalin Zheng},
  title = {{Automated Detection of Vessel Abnormalities on Fluorescein Angiogram in Malarial Retinopathy}},
  booktitle = {Scientific},
  year = {2015},
  doi = {10.1038/srep11154},
}
Figure