Spatial Correlation Module for Classification of Ocular Diseases in Diabetic Retinopathy Using Color Fundus Images

Abstract

Early identification and therapy of ocular diseases (ODs) are crucial in avoiding irreversible vision loss. Colour imaging of the fundus (CFI) is an economical and trustworthy screening tool. However, automatic and thorough diagnostic tools are required since early OD symptoms are usually modest. The traditional wisdom suggests treating the eyes independently and relying only on image-level diagnostics, without considering interocular correlation data. Additionally, these techniques typically only target a single or a limited number of ODs simultaneously. This research presents a novel classification model called PLML_ODs. It includes patient-level multi-label OD data. Our technique integrates patient-level diagnosis, notably in diabetic retinopathy, by combining bilateral ocular and multi-label ODs classification. A feature correlation SCNet, a classification score generator, and a feature extracting backbone based on convolutional neural networks (CNN) named DenseNet-169 comprise the PLML_ODs system. The DenseNet-169 model obtains two sets of features using on both sides CFI. The SCNet then captures a correlation connecting the two collections of features pixel-wise. In order to get an embodiment at the patient level, the attributes are integrated after analysis. Using this representation, the ODs classification process is carried out. We evaluate PLML_ODs's classification performance to that of other baseline techniques using a publicly available dataset and an enveloping margin loss.