Deep multiple instance learning for automatic glaucoma prevention and auto-annotation using color fundus photography

In the area of ophthalmology, glaucoma affects an increasing number of people. It is a major cause of blindness. Early detection prevents severe ocular complications such as glaucoma, cystoid macular edema, or diabetic proliferative retinopathy. Intelligent systems are proven to be beneficial for the assessment of glaucoma. In this paper, we describe an approach to automate the diagnosis of glaucoma disease, based on color funds photography using deep learning. The setup of the proposed framework is ordered as follows: The bidimensional empirical mode decomposition (BEMD) algorithm is applied to decompose the ROI to components (BIMFs + residue). CNN architecture VGG19 is implemented to extract features from decomposed BEMD components. The features obtained are the input parameters of the implemented classifier based on full connect layers and softmax. To train the built model, we have used the public dataset RIM-ONE DL. To test our models, we have used a part of RIM-ONE DL and REFUGE. The average obtained sensitivity, specificity, accuracy and AUC rates are, respectively, 99.14%, 99.19%, 99.13%, 99.09% and 99.17%, 99.24%, 99.20%, 99.18% in RIM-ONE DL and REFUGE dataset. The experimental results obtained from different datasets demonstrate the efficiency and robustness of the proposed approach. A comparison with some recent previous work in the literature has shown a significant advancement in our proposal.