COMPARATIVE ANALYSIS OF METHODS FOR BLOOD VESSEL DETECTION IN RETINAL IMAGES

Radovan Turović ； Gorana Gojić ； Dinu Dragan ； Veljko Petrović ； Dusan Gajić ； Ana Oros

《IETI Transactions on Ergonomics and Safety》 5卷1期 (2021/05) Pp. 24-35

https://doi.org/10.6722/TES.202105_5(1).0004

摘要

Images of the back part of the eye, also known as retinal images, are the basis for the diagnosis of many systemic and eye diseases such as glaucoma, diabetic retinopathy, and retinopathy of prematurity. Disease indicators may be found by observing the blood vessel network in the retinal image. A failure of an ophthalmologist to correctly identify a disease, due to fatigue or a low-quality image, may lead to severe health damage. To address this problem, many methods for automatic vessel detection in retinal images have been proposed. Among those, machine learning approaches based on convolutional neural networks have proven to yield the best results. Often, these methods require some sort of input retinal image preprocessing, such as transformation to grayscale, to emphasize blood vessels on images and reach their full potential. In this paper, we employ a subset of general-purpose algorithms for edge detection to produce retinal images with an emphasized retinal blood vessel network, which can be used for convolutional neural network blood vessel detection training. We test Canny, Sobel, Scharr, and Hollisticaly-Nested Edge Detection algorithms on the DRIVE dataset. Resulting images produced by these four algorithms are evaluated by an experienced ophthalmologist. Each image was graded and the time required to make the decision was measured. The ophthalmologist (who operated under double-blind test conditions) was later interviewed and qualitative data was collected. The data was then analyzed showing a clear win for the Sobel algorithm which, according to the post-test interview, preserves more fine detail.

關鍵字

Computer vision ； retinal images ； blood vessel ； edge detection