An Automated Exudate Detection Scheme Supporting Diabetic Retinopathy Screening Using Spatial-Spectral-statistical Feature Maps

An automated Diabetic Retinopathy (DR) introspection scheme for early detection of retinopathy signs is realized in this paper. Presence of exudates in retina is considered as an early sign of DR. Therefore, the proposed methodology aims at detection of exudates by transforming the acquired fundus image into a high-dimensional feature map labelled as Spatial-Spectral-Statistical (SSS) feature map that uniquely represents the individual image pixels using a novel characterization scheme. At the onset, a slightly novel pre-processing scheme is fused into the mechanism to address the non-uniform illumination issues present in images. Later, separate feature characterizers and descriptors pertaining to the diverse domains are engaged for extraction of the different features from the input fundus image. These distinct features are then blended to yield the feature map representing the given image. Then a supervised classifier categorizes these features and finally aids in deciding the presence or absence of exudates for the given input. Extensive investigation and relative comparisons performed on publicly available dataset namely DIARETDB0, DIARETDB1 and MESSIDOR demonstrate a consistent average classification accuracy of 97.99%, an attribute owed to the unique feature aggregation scheme that also, makes the methodology robust under different imaging problems.