Optic disc detection in the presence of strong technical artifacts. (August 2019)

• Record Type: Journal Article
• Authors: Dietter, Johannes; Haq, Wadood; Ivanov, Iliya V.; Norrenberg, Lars A.; Völker, Michael; Dynowski, Marek; Röck, Daniel; Ziemssen, Focke; Leitritz, Martin A.; Ueffing, Marius
• Is Part Of: Biomedical signal processing and control. Volume 53(2019)
• Publication Date: 2019-08
• Subjects: Optic disc detection -- Optic disc segmentation -- Technical artifacts; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2019.04.012
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Optic disc (OD) detection and segmentation in the presence of strong technical artefacts. Two-stage approach of OD-detection and segmentation advancing pre-existing methods. OD detection via analysis of vessel orientation, vessel thickness and brightness information. Improved OD segmentation by loess-smoothing of the OD-radius depending on circular angle. Improved detection rate of 95.9% (5052 Images, public + private), overlap ratio of 88.8%. Abstract: The inspection of retinal fundus pictures taken by a fundus camera is one of the key procedures for diagnosing a wide range of diseases like diabetic retinopathy, hypertension or hypercranial pressure. The detection of the optic disc (OD) is of particular relevance since changes of the OD itself may indicate specific diseases. Moreover, the OD serves as a landmark for retinal image analysis. Here, we present a method to detect and segment the OD that can cope with strong technical artifacts. Conceptually building on two published methods, we developed a two-stage approach to localize and then segment the border of the OD. First, we use vessel orientation and brightness to determine the center of the OD. Second, we modify a score function from literature whose maximum indicates the pixel that possesses the best OD-border like structure around it. Using six publicly available and three in-house databases, we analyzed a total of 5052 retinal images, resulting in an average detection rate of 95.9%. Taking a selection of 61 pictures of one of our in-house datasets, we achieved an average overlap ratio of 88.8% between the OD marked by an expert and the OD determined through our algorithm.