Opthalmic Image analysis and machine learning for eye disease detection

National Library Of Medicine

Investigators

Abstract

Glaucoma. In a collaboration with the National Eye Institute (NEI), we experimented with several Convolutional Neural Network-based (CNN) Deep Learning (DL) models to detect regions of interest (ROI) in ophthalmic fundus images. Localizing the ROI is necessary, as a first step, to accurately detect optic cup and disc areas to compute the likelihood of glaucoma. Importantly, we found that results were improved by using blood vessels as features. These blood vessels were extracted by an algorithm we developed using image analysis techniques and also experimenting with the SegNet DL technique. Several open image data sets were used for the analysis including 600 images from the MESSIDOR set (original images and the ones embedded with blood vessels) for training the DL algorithm. For testing we used 1800 images from MESSIDOR, 6 other open source datasets and the AREDS dataset from NEI. The results showed that using the images embedded with blood vessels yielded the best performance overall: 99.33% accuracy for MESSIDOR, 97.74% for Open Sources, and 90.25% for AREDS (from NEI). These results are 2 to 3% better than previous experiments done without blood vessels as features. The next step is to use the ROI to detect cup and disc areas. We also developed a CNN-based technique to classify age-related macular degeneration (AMD) disease into levels of severity. NEI categorizes the images in six stages of increasing severity (1, 2, 3a, 3b, 4a, and 4b). We collected 5,898 images and trained a CNN (VGGNet16) DL model to classify images into two classes: Stages 1 and 2 (low) vs. Stages 3 and 4 (severe). 4,719 images were used for training and 1,179 images were used for testing. The test results showed an accuracy of only 85.58%, since 55 images from Categories 1 and 2 and 115 images from Categories 3 and 4 were misrecognized. Research toward reducing these errors is ongoing.