Assessment of Choroidal Neovascular Membranes Secondary to Age-Related Macular Degeneration Using Feature Selection

Abstract

In this paper we investigate whether geometric features can potentially distinguish between classic lesions, occult lesions and areas of blood leakage in fundus fluorescein images. Initially 37 features were identified following consultation with ophthalmic graders. To test this concept, four test images, two classic and two occult, were analysed using image processing algorithms for region of interest identification. Features were then derived; six were global parameters; one was lesion area, and the remaining thirty were local measurements based on cross-sectional slice of the lesion. Each feature was assessed for its potential to discriminate between lesions types. A subset of 16 features was further investigated in an attempt to classify lesions as classic or occult and to identify areas of blood leakage. Whilst some of the features selected show promise, the best position for each cross-sectional slice has still to be determined. We conclude that quantitative measures should be applied for classification in this domain. Validation of this approach requires a much larger set of labeled images, selection of complementary features and the use automated feature ranking.