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 applied several Convolutional Neural Network (CNN) Deep Learning (DL) models to detect regions of interest (ROI) on ophthalmic fundus images to segment optic disc and cup. We used three different DL models: modified LeNet, Faster Region-based CNN (Faster RCNN), and RetinaNet. For the MESSIDOR dataset containing 1,200 fundus images, the modified LeNet showed 0.9950 accuracy in ROI detection, but needed long processing time. Faster RCNN improved the processing time (about 1.5 second per image) and resulted in an accuracy of 0.9900. RetinaNet showed the best performance with 1.0000 accuracy and the best processing time (0.2 second per image). Optic disc and cup segmentation from the ROI in fundus images is the second step. We used Fully Convolutional Network (FCN) models. In the case of optic disc, we trained two different FCNs: FCN2 (FCN for two classes) and FCNM (FCN for multi classes). The RIGA dataset, containing 750 fundus images with annotations by six ophthalmologists from three different datasets (MESSIDOR, Bin Rushed, and Megrabi), was used for the test. The results showed that FCN2 had the best performance with 0.9430 Jaccard Index, 0.9702 F-measure, and 0.9889 Accuracy. In the case of cup, we used the two FCNs (FCN2 and FCNM) and two different ROIs (original ROI and ROI masked by optic disc (masked ROI)) for the test. The FCN2 using the masked ROI showed the best performance with 0.8037 Jaccard Index, 0.8873 F-measure, and 0.9882 Accuracy. Our results showed over 1% better performance in optic disc and 2% better performance in cup than other existing methods. The next step will be to apply these techniques and estimate optic disc and cup ratio on the RIGA dataset and AREDS dataset from NEI. AMD: NEIs AREDS dataset contains fundus images of 4,757 patients with AMD and no AMD. NEI categorizes the fundus images into twelve AMD severity levels. We developed CNN-based techniques to classify fundus images into the twelve severity levels. 84,448 images were collected for training and 13,181 images were for testing from the dataset. Two CNN models, VGG16 and ResNet101, were used for training. The test results showed that VGG16 had 0.6267 accuracy and ResNet101 had 0.6628 accuracy. We plan to use several different CNN models and different loss functions to improve the accuracy. Uveitis: We also experimented with FCN models to segment leakage and blood vessel areas in fluorescein angiography (FA) images. Since uveitis is a rare disease, there were not many images to use. We received 90 images from NEI, among which 50 were annotated. Five-fold cross-validation, image cropping, and image augmentation techniques were used for the test. We trained two FCNs to segment leakage and blood vessel in the FA images. In the case of leakage, the FCN showed 0.3157 Jaccard index, 0.6241 sensitivity, 0.9669 specificity, and 0.9611 accuracy. In the case of blood vessel, the FCN showed 0.4022 Jaccard index, 0.7168 sensitivity, 0.9824 specificity, and 0.9766 accuracy. The next step is to improve segmentation accuracy by using more FA images and different FCN models.