CDLRS: Collaborative Deep Learning Model with Joint Regression and Segmentation for Automatic Fovea Localization

With the development of information technology, eyes are easily overworked for modern people, which increases the burden of ophthalmologists. This leads to the urgent need of the computer-aided early screening system for vision examination, where the color fundus photography (CFP) is the most economical and noninvasive fundus examination of ophthalmology. The macula, whose center (i.e., fovea) is the most sensitive part of vision, is an important area in fundus images since lesions on it often lead to decreased vision. As macula is usually difficult to identify in a fundus image, automated methods for fovea localization can help a doctor or a screening system quickly determine whether there are macular lesions. However, most localization methods usually can not give realistic locations for fovea with acceptable biases in a large-scale fundus image. To address this issue, we proposed a two-stage framework for accurate fovea localization, where the first stage resorts traditional image processing to roughly find a candidate region of the macula in each fundus image while the second stage resorts a collaborative neural network to obtain a finer location on the candidate region. Experimental results on the dataset of REFUGE2 Challenge suggest that our algorithms can localize fovea accurately and achieve advanced performance, which is potentially useful in practice.