Zheng Y, Zhao Y, Chen X, Gao D, Bridge J, Zhu W, Williams B. Fully automatic localisation of the optic disc using YOLO in colour fundus photographs. Investigative Ophthalmology & Visual Science. 2019 Aug 1;60(11):PB038-.
https://iovs.arvojournals.org/article.aspx?articleid=2748122
Abstract
Purpose
Detection of the optic disc (OD) is important for the management of eye disease. Knowledge of the OD is considered essential for the diagnosis and screening of many retinal diseases, with the OD centre often regarded as a reference point for locating other retinal structures in the assessment of colour fundus photographs. The purpose of this work is to develop a fully automatic method for the localisation of the OD from colour fundus photographs.
Methods
We propose a fully automatic method for OD localisation in colour fundus photographs. The proposed method is based on a YOLO (you only look once) network architecture, which allows the simultaneous localisation of an object of interest and detection of the bounding box. 2708 images from 7 publically available datasets were used along with corresponding manual OD annotations from expert graders. We trained the network on 1508 colour fundus images from 6 publically available datasets (DRIONS: 110; DRISHTI: 101; ONHSD: 88; ORIGA: 650; REFUGE: 400) and tested on the 1200 images from the 7th dataset (MESSIDOR). Localisation accuracy was evaluated in terms of ¼, ½ and one OD radius (ODR).
Results
State-of-the-art detection results were achieved, demonstrating the excellent performance of the method and robustness of the detection. The detected OD centres in 1149 images out of 1200 images (95.75%) were within ¼ ODR of the annotated centre. The detected OD centres were within ½ and one ODR for 1199 images (99.9%). Importantly, these results have been achieved in the external validation of our model with data from a different population. Six localisation examples are shown in the figure.
Conclusions
A reliable and accurate automation method was proposed for the localisation of the OD and validated on external data, achieving state-of-the-art results. This can save considerable amounts of time, improving disease management and diagnostic potential, and paving the way for complete, fully automated systems to be realised for diagnosing eye disease.
