A Multi-Layer stacked ensemble classifier model for improved classification accuracy of Maculopathy gradation

Abstract

Diabetic Maculopathy (DME) is the serious impediments of diabetes, which may cause permanent blindness unless timely detected. Vision impairment because of diabetes is substantially avoidable with well-timed screening and intervention at primary stages. Presence of most primitive and distinctive signs on the retinal surface is micro-aneurysm and haemorrhage, signify as dark spots and hard and soft exudates signifies as bright lesions. Hence, recognition of all these bright lesions is the first step of automated recognition of DME. In this paper, we present a multi class, multi-layer stacked ensemble classifier-based model with four base learners and one meta-learner for improved exudates (EXs) classification accuracy and maculopathy gradation system. The proposed system involves pre-processing, Scale-Space Extrema Detection(SSED) based extraction of clinically significant bright lesions, shape, colour, intensity, and statistical functions-based feature set creation, Minimum Redundancy-Maximum Relevance (mRMR) feature selection, stacking classifier with Bayesian optimization (BO) for hyper-parameter tuning and severity gradation. Information of location of all types of exudates is accounted for to provide the level of severity of DME. At both the image and lesions levels, the proposed system's quantitative assessment is carried out utilising publicly available databases. When compared to other state-of-the-art methodologies, our system's results have achieved competitive performance in three and two class exudates classification and DME gradation.