Automatic Digital Analysis System to Grade Diabetic Retinopathy by Integrated Stacking Model Concept

Ophthalmic diagnosis is based primarily on visual information (photographs of the retina), and the recent availability of digital fundus images allows such quantization of parameters or classification to be derived from computerized image processing.Thus, an Automated Digital Analysis System (ADAS) is designed for grading Diabetic Retinopathy (DR) in this work.The Integrated Stacking Model (ISM) concept is employed to design a single large multi-headed deep learning model.Each sub-model has three layers with a predefined number of 33 convolution filters (128, 256 and 512) and a pooling layer to abstract deep features.The outputs of sub-models are fed to the meta-learner for grading the DR.The performance of the ADAS for grading DR is evaluated using MESSIDOR-1 and Kaggle datasets.For the images in MESSIDOR-1, the proposed system is considered a four-class problem, and it is a five-class problem for Kaggle dataset images.Results show that the proposed ISM-DR classification system provides promising results with an average classification accuracy of 99.2% (four-class) and 99.1% (five-class) using MESSIDOR-1 and Kaggle dataset images respectively when stacking five sub-models.A clear trend can be seen from the experimental results towards better performance when more sub-models are stacked.