Automated Segmentation of Foveal Avascular Zone in Fluorescein Angiography in a Clinical Setting

Abstract

Purpose: : To investigate the effectiveness in a clinical setting of an automated approach to segment the foveal avascular zone (FAZ) on fluorescein angiography (FA) images.

Methods: : FAZ refers to the capillary-free area of the central macula. Its appearance can be used to indicate differing stages of diabetic maculopathy. FA sequence was acquired on HRA2 with high resolution setting after injection of 2-3 ml 20% dye within 2-3 seconds. For each sequence a single frame acquired in the transit phase was analysed: A sub-image containing the FAZ was cropped from the original image and was then smoothed by Gaussian kernel (σ=1.5). An initial contour manually placed inside the FAZ of the smoothed image was iteratively moved by the segmentation program towards the FAZ boundary guided by the gradient of the smoothed image. Figure (a) and (b) illustrated a sub-image and its segmentation result respectively.

Results: : Images from 16 patients were studied. Visual inspection showed tests on 9 images produced satisfactory accuracy while tests on the rest 7 images failed due to poor image quality (Figure (c)). In order to test the repeatability an average overlap ratio R_a (intersection/union) over 10 tests with different initialisations were obtained for each of the images. For the former 9 images good repeatability was observed (mean (±sd) of R_a: 97.1% (±3.1%) with a range of 91.2% to 99.6%). For the rest 7 images, repeatability suffered (mean (±sd) of R_a: 49.6% (±31.4%) with a range of 92.6% to 18.3%).

Conclusions: : The proposed segmentation approach showed potential for automated segmentation of FAZ. However in a clinical setting its use requires further optimisation of image acquisition and processing.

Keywords: imaging/image analysis: clinical • shape and contour • diabetic retinopathy