Abstract

This paper introduces SAGE--an algorithm that uses the spatial clustering of objects to enhance their classification. It assumes that discrete objects can be identified and classified based on their individual appearance, and further that they tend to appear in spatial clusters (for example, circinate exudates). The algorithm builds spatial distribution maps for objects and confounds for a given image, and adjusts individual object confidence levels to reflect their spatial clustering. SAGE may be combined with a wide range of object identification and classification methods; we demonstrate it using a Multi-Layered Perceptron (MLP) Neural Network and a Support Vector Machine (SVM) classifier types for both dark and bright retinal lesions. Using ROC analysis SAGE improves classifier performance as much as 83.