RetinaSC: Retinal Semantic Communication Framework for Diabetic Retinopathy Detection

This research-based project explores RetinaSC, a lightweight semantic communication framework designed for diabetic retinopathy (DR) grading in telemedicine applications. Traditional image transmission over wireless channels often suffers from high bandwidth demands and latency. To address this, the framework utilizes task-oriented semantic communication where only critical diagnostic features are extracted and transmitted, rather than raw pixel data. The system evaluates multiple semantic encoder architectures, including ResNet34, U-Net, MobileNetV2, and EfficientNet-B0, trained on the APTOS2019 fundus image dataset. Additionally, transfer learning is applied to adapt the trained models to the IDRiD dataset. The paper reports that task-oriented semantic systems achieve a weighted F1 score exceeding 0.8 using ResNet34 and EfficientNet-B0 across various

channel conditions. Notably, MobileNetV2 offers an optimal balance between computational complexity, throughput, and diagnostic accuracy, achieving a weighted F1 score of 0.76 to 0.78 with an end-to-end latency of 8.162 ms. This framework provides a solid foundation for M.Tech students researching joint source-channel coding, deep learning-based wireless transmission, and resource-constrained medical IoT systems.

Message: Hi FE, I need help with “RetinaSC: Retinal Semantic Communication Framework for Diabetic Retinopathy Detection” in “Electronics & Communication Engineering”