RetinaSC: Retinal Semantic Communication Framework for Diabetic Retinopathy Detection
This paper presents a lightweight semantic com- munication framework for diabetic retinopathy (DR) grading from fundus images pertaining specifically to the APTOS2019 dataset. The framework includes semantic encoder designed using ResNet34, U-Net, MobileNetV2, and EfficientNet-B0 architectures. Further, the trained models are used to train the same framework for IDRiD dataset using transfer learning approach. Such a framework will enable the telemedicine sector in the early detection of DR and prevent total blindness. The results reveal that for APTOS, task-oriented semantic systems can achieve satisfactory robust performance (more than 0.8 weighted F1 score for ResNet34 and EfficientNet-B0) across different channels. In particular, considering the trade-off between complexity, latency, throughput, and F1 score, MobileNetV2 is the best model with weighted F1 score of 0.76-0.78 for both task and reconstruction- oriented modules and end-to-end latency of only 8.162 ms.