Automated Segmentation of Capillary Non-perfusion (CNP) Regions in Fundus Fluorescein Angiograms (FA) Using a Texture-based Approach

Abstract

 

Purpose:

 

To describe and evaluate a technique for automated detection of CNP regions using a texture-based technique and explore applicability in ischemic retinal disease.

 

Methods:

 

The proposed segmentation framework consists of four major steps and was tested on a collection of fundus images from children with malarial retinopathy. Firstly a mask of the FA image under consideration (Fig. A) was generated through thresholding and morphological operations in order to restrict the region of interest to the camera aperture. Then the image underwent correction for uneven background illumination by applying the top-hat filter with a structural element size of 100 empirically chosen for the optimal correction. Then the pre-processed image was segmented into CNP and non-CNP, using a variational texture segmentation model coupled with a fast optimisation strategy. The segmentation result was further refined by removing the fovea (if it appeared in the image) in an interactive manner to reduce CNP false positives and the final result is shown in Fig. C. FA images with CNPs taken with a fundus camera (TRC-50 EX, Topcon, Tokyo, Japan) were collected. All images were examined by a retinal specialist, and the CNP regions were delineated manually as ground truth for evaluation purposes (Fig. C). The performance of the program was measured (the number of pixels correctly segmented against the total number of pixels).

 

Results:

 

The proposed technique was applied to a collection of 40 FA images with size of 3008x1960 pixels from eyes diagnosed with malarial retinopathy, and an average accuracy of 81.0% was achieved.

 

Conclusions:

 

CNP is a hallmark feature of ischaemic retinal diseases. The results indicate that this technique shows promise for measuring CNP in FA images and suggest it has the potential to become a powerful tool in the management of ischaemic retinal diseases.  

 

Keywords: image processing • imaging/image analysis: clinical • retina