Classification of age-related macular degeneration using very deep learning neural network based on transfer learning

Abstract Background: Detecting and classifying the age-related macular degeneration (AMD) at the elderly people is one of the main concerns of the national public health program in Thailand. Within this research purview, accurate and sensitive performance is highly desirable, and a variety of proposed models have been developed toward this end. Deep learning neural networks have recently been shown to have advantage over existing approaches for the task of classifying the level of eye disease. Despite of these advances, there is still significant potential for development, regarding model classification accuracy, and sensitive values. Results: Six very deep learning neural networks (DLNN), named as InceptionV3, ResNet152V2, DenseNet201, EfficientNetB7, InceptionResNetV2, and NASNetLarge are proposed for training to detect and classify the AMD disease from fundus images. The training process with AMD images is implemented based on the transfer learning technique through Google Colab Pro platform. The experiments showed that, for 3-classes AMD classification (Normal, Dry AMD, and Wet AMD), the overall classification accuracy of DenseNet201 is highest and is about 88% with the testing dataset. Furthermore, for the inferred 2-class AMD classification (Normal vs. AMD), the most accuracy values obtained from EfficientNetB7 and InceptionResNetV2 about 93.5% and 95.06%, respectively. Conclusions: The experimental results have shown that, after applying transfer learning technique, the DenseNet201 model had higher accuracy performance in 3-class AMD classification of fundus images than other very deep learning neural networks and other state-of-the-art proposed deep learning models in the literature. Furthermore, the trained EfficientNetB7 and InceptionResNetV2 showed the advantages for the 2-class AMD classification over other existing models.