Multiocular disease detection using a generic framework based on handcrafted and deep learned feature analysis
M Raveenthini, R Lavanya
Abstract
Diabetic retinopathy (DR) and Glaucoma are two major ocular diseases that lead to vision impairment if not detected and treated promptly. Manual detection and diagnosis of these diseases is a laborious and time-consuming process. Computer aided diagnosis (CAD) systems can serve to assist physicians in manual diagnosis. This work aims to develop a generic multiocular CAD system for DR and glaucoma diagnosis, which could serve as a boon in a large scale screening scenario by reducing the time and manpower involved. To this end, a segmentation-independent approach is employed that eliminates the need for individual diagnostic systems each involving a set of disease-specific algorithms for localisation and analysis of regions of interest (ROIs). Further, the proposed approach alleviates the segmentation inaccuracies attributed to image quality and anatomical factors, which have a cascaded effect on the classification performance. Specifically, a machine learning (ML) model based on random forest (RF) classifier and a pool of non-linear features including higher order spectra (HOS), entropy and fractal features, was developed. An ensemble of this ML model and convolutional neural network (CNN)-based deep learning (DL) model was further constructed using the ‘sum rule’ for decision fusion. The proposed ensemble model resulted in accuracy, sensitivity and specificity as high as 98.08%, 98.37% and 99.07% respectively, for three-class classification to categorize samples as normal, DR or glaucoma.