Diverse lesion detection from retinal images by subspace learning over normal samples

Abstract Lesion detection from retinal images is an important topic in the retinal image analysis. Many computer-aided detection techniques have been developed for detecting retinal lesions. However, these techniques are mainly used to detect specific lesion types from retinal images. They cannot be applied to detect diverse types of lesions from retinal images, which is a challenging task because lesion number and types in retinal images are generally unknown in advance, and different lesions may exhibit diverse properties in shapes, sizes, colors, textures and positions. Inspired by the doctors’ visual diagnostic mode, this paper proposes a novel computational framework to detect various types of lesions from retinal images. In this framework, many healthy fundus images are collected to act as ”doctors’ detection experience”, and local visual properties of lesions are used to distinguish true positives from false positives. A specific subspace is learned from the collected normal set and acts as a specific structural filter, by which various lesions in a retinal image can be filtered out while other normal regions keep little changes. By computing the difference image between a target image and its filtered image, different types of lesion candidates can be separated from the image. Furthermore, based on local visual context properties of lesions, the true lesions are identified from the lesion candidates. Extensive experiments have shown that the proposed method can more effectively detect diverse lesions from retinal images compared with related methods.