Automated detection and quantification of microaneurysms in fluorescein angiograms

Abstract

Fluorescein angiograms from diabetics were digitised for analysis using digital image-processing techniques. Computer algorithms were written to detect and count microaneurysms present in the images. The accuracy, speed and reproducibility of the technique were assessed and compared with those of manual counts made by clinicians from both digitised and analogue images. Free-response ROC (receiver operating characteristic) curves were used to assess the performance of both the clinicians and the computer by comparing the results with "gold standards" compiled from prints of the original fluorescein angiograms. The computer performed as well as the clinicians when the latter were analysing the digitised images (512 x 512 pixel resolution), but only when one image was acquired at 4 times this resolution did the computer's performance match that of the clinicians analysing the analogue image. The automated technique was more reproducible than the manual method.

Automated detection of microaneurysms using scale-adapted blob analysis and semi-supervised learning

“…This performance is achieved with few manually annotated training samples. Moreover, the increased in performance obtained by using scale-adapted (variable σ) features against fixed scale approach asserts the importance of [7] Top-hat transform 12% 20.3 Abdelazeem [17] Circular Hough-transform 28% 505.85 Walter et al [10] Diameter closing 36% 154.42 Zhang et al [16] Multiple-Gaussian mask 33% 328.3 Lazar et al [14,15] Cross Table 4: FROC results on of retinopathy challenge testing dataset at low false positives per image.…”

“…Background estimation and subtraction methods are the most commonly used techniques to correct shading effects [7]. This method uses the green channel of the image to estimate the "background" image using a large median filter of size 25 × 25 and sub-tract it from the green channel resulting in the shade corrected image.…”

“…A mathematical morphology based candidate microaneurysm extraction method is proposed by Spencer et al [7] and Frame et al [4]. The proposed approach removes bright-lesions by setting all positive pixel values of shade corrected image to zero.…”

Detección automática de microaneurismas en retinografías

“…Aunque en los últimos años han surgido publicaciones sobre la detección automática de microaneurismas [3][4][5][6][7] , son pocas las que se centran en la detección en imágenes de retinografía [8][9][10][11] .…”

“…Un número importante de publicaciones en el campo de la ingeniería abordan este problema 2 . La mayoría de ellas se centra en la detección de MA en angiografías de fluoresceína [3][4][5][6][7] . Øien y Osnes fueron los primeros en detectar MA en imáge-nes en color 8 .…”