Development of Lightweight Deep-Learning Scheme to Classify Retinal OCT into Normal/DME Class

Eye is a vital sensory organ, which is responsible to convert the light into information. The abnormality in eye leads to a variety of vision issues and early diagnosis and treatment is necessary. This research proposes a Lightweight Deep-Learning Technique (LDLT) to examine the retinal Optical Coherence Tomography (OCT) images. The stages of this research includes; (i) data collection and resizing, (ii) feature mining using the chosen LDLT, and (iii) binary classification using five-fold cross validation and performance confirmation. In this study, the retinal OCTs of normal (1000 images) and Diabetic Macular Edema (DME) (1000 images) are considered for the examination with MobileNets (MNs) and NASNets (NNs). The classification of the OCTs is performed using the deep-features and the performance metrics computed during this task is then evaluated to confirm the merit of proposed tool. The experimental outcome of this study confirms that the MobileNetV2 (MNV2) helped to achieve a detection accuracy of >97% with the chosen data. In future, ensemble of features and fused deep-features are to be included to improve the DME detection accuracy.