OPTIMIZED DEEP LEARNING MODEL FOR OCT-BASED RETINAL DISEASE ANALYSIS

Optical coherence tomography is an effective non-invasive imaging modality used for producing high-resolution images in cross-sections that enables easy identification of various retinal diseases, such as Choroidal Neovascularization (CNV), Drusen, and Diabetic Macular Edema (DME). Early detection of such diseases allows avoiding loss of vision. Retinal diseases detection is traditionally based on extracting handcrafted features by utilizing techniques such as GLCM, LBP, and edge detection followed by classification by algorithms such as SVM, k-NN, Decision Trees, and Naïve Bayes. Such conventional models suffer from their reliance on expert-defined features, inability to extract complex retinal information, lack of ability to adapt to changes in image quality, and inefficiency.The developed system solves those limitations through the use of pre-trained Xception CNN model, optimized convolutional and fully connected layers, and cache-based feature optimization allowing to classify OCT retinal images as belonging to CNV, Drusen, DME, or Normal classes. Deep learning technology allows achieving high accuracy, minimal computational overheads, and applicability for big data analysis, ensuring that the ophthalmologist will be provided with highly-efficient clinical decision-making support tool for early detection of retinal diseases.