Region of Interest Detection in Fundus Images Using Deep Learning and Blood Vessel Information

Abstract

Ophthalmologists use the optic disc to cup ratio as one of the factors to diagnose glaucoma. The region of interest (ROI) for glaucoma in fundus images is the area that locates optic disc and cup in the center. Therefore, ROI detection is used as a preprocessing step for automatic detection of optic disc and cup areas. This paper proposes an automated method to detect ROI using deep learning. Convolutional Neural Networks (CNNs) are used to classify ROI and non-ROI images. The structure of our CNNs is composed of two convolutional layers, two Max Pooling layers, two fully connected layers, and one output layer. We train two CNNs using fundus images from the MESSIDOR dataset, a public dataset containing 1,200 fundus images. In addition, we estimate blood vessels from the images and use the images embedded with the blood vessels to train two other CNNs. The proposed method moves a window in the horizontal and vertical directions in each fundus image, estimates a probability of each window using the CNNs, and selects the window with the highest probability as ROI. The experimental results are promising. The best-performing CNN from the first CNN group shows over 0.99 accuracy for the MESSIDOR dataset and over 0.93 accuracy for five other public fundus image datasets. The best CNN from the second CNN group shows more robust results: over 0.99 accuracy for the MESSIDOR dataset and over 0.97 accuracy for the five other image datasets.