A Comparative Investigation of Transfer Learning Frameworks Using OCT Pictures for Retinal Disorder Identification

Abstract

Determining indicators or retinal wounds is essential for accurate diagnosis and retinal disorders grading. To view the retinal microarchitecture and easily screen for anomalies, optical coherence tomography (OCT) images are utilized. Numerous studies have already tried to use OCT to overcome that issue. Throughout this study, we describe an OCT image-based transfer learning (TL) approach for the identification of four retinal diseases. This study compares four distinct models with one another. A MobileNetV2 model’s detection accuracy on the test set is 100%; an InceptionNetV3 model’s is 99.9%; an EfficientNet model’s is 99.38%; and a DenseNet model’s is 99.79%. The InceptionNetV3 model approaches the highest accuracy, while MobileNetV2 model achieves the maximum accuracy. The suggested method may influence the development of a tool for automatically identifying retinal disorders. The promising suggested architecture’s qualitative assessments and quantitative outcomes through creating a confusion matrix demonstrate how the suggested methodology can be utilized in healthcare settings as a diagnostic tool to assist medical professionals in making more accurate diagnoses.

Keywords

• Transfer learning
• OCT
• deep learning
• classification
• image enhancement
• retinal disease