Automated Detection of Vessel Abnormalities on Fluorescein Angiogram in Malarial Retinopathy

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.

Figure 1. Two example fluorescein angiography images illustrating the appearances of IVFDs.

Vessels with IVFD are shown by single arrows. Vessels without IVFD, in the same image, are shown by double arrows.(a) Example 1: Intensity of mature parasitized red blood cells in vessels with IVFD is significantly different from normal vessels. (b) Example 2: Edges of vessels with IVFDs become unsmooth, the diameter is changed dramatically when compared to normal vessels. The images on the right are the zoom-in view of the regions enclosed by the green box within the original image on the left respectively.

Figure 2. Comparison of abnormal vessel detection between the automated framework and manual annotations (abnormal vessels are highlighted in red and normal vessels in green).

(a) Three example fluorescein angiography images. The inset in each image shows the zoom-in view of the region enclosed by the green box within that image. (b) Detection results on all the segmented vessels by the framework. (c) Detection results on the gradable vessels only by the human observers. (d) Manual annotations from a consensus of two observers was used as the reference standard.

Figure 3. Evaluation results in terms of sensitivity, specificity, accuracy, and area under the curve (AUC), under different combinations of decision trees and cost coefficients.

(a) Results using 500 trees; (b) Results using 2000 trees; (c) Results using 5000 trees. On each plot, from left to right are the results with cost coefficient of 6, 7, and 8 respectively.

Figure 4. Illustration of the detection performance of the framework on vessels with different sizes (Left column: Manual annotations. Right column: Automated annotations).

From top to bottom, (a) Results on large vessels. (b) Results on small vessels. (c) Results on peri-capillary vessels.

Figure 5. Illustration of vessel segmentation and geometric analysis results.

(a) Two example fluorescein angiography images. (b) Vessel segmentation results. (c) Vessel segments after removing branch pixels from images in the second row. In (b) and (c) pixels in white denote vessels.

Figure 6. Illustration of saliency maps.

(a) Two example fluorescein angiography images. (b) Saliency maps of each individual vessel segments. (c) Vessels are divided into salient and non-salient regions after applying thresholding process to images in the second row respectively. Blue colour indicates the most salient regions while red colour shows the least salient regions for images in the second and third rows. The inset in each image shows the zoom-in view of the region enclosed by the green box within that image.