Leveraging fundus images for on device eye disease diagnosis with AI powered lightweight software hardware framework

Vision loss due to illness can result from various medical conditions that affect the eyes. Advanced devices like OCT and ultra-widefield retinal cameras are expensive, making them less accessible in resource-limited settings. While eye image capture devices have transformed disease detection, they require expert analysis, and also variability between observers can lead to inconsistent diagnoses. Many devices do not seamlessly integrate with electronic health records (EHR), complicating workflow and data management. In the detection and diagnosis of eye disorders, machine learning (ML) has emerged as a vital technique. Deep learning algorithms based on individual eye diseases have been created to diagnose them. These models still have a number of drawbacks, including complexity of computation, low precision brought on by class imbalance, and increased time consumption. In order to support ophthalmologists, this study suggests an effective software-hardware framework with edge AI integration that uses a refined EfficientNetB0 model with spatial attention CNN block for rapid and accurate identification of eight types of eye illnesses using fundus images. While using pre-trained models and heatmaps is common, combining EfficientNet-B0 as a feature extractor with ImageNet-trained models for downstream classification introduces a novel hybrid pipeline. The architecture of the developed efficient net B0 model is optimized for the low-cost hardware and the work is directed to develop a smart, low latency, cost-effective device for use in underprivileged areas. Publicly accessible datasets from various sources, including 13,300 fundus and private hospital images, are used in the experimental work. The EfficientNetB0 model outperformed all other models with an accuracy of 96.2% for multi-label tasks, followed by Unet with 88% and DenseNet169 with 91%. The key aspects of the process are model optimization for edge devices, efficient preprocessing, and leveraging TensorFlow Lite's performance features for real-time disease detection with an average inference time of 5 s per image on the Raspberry Pi 5 platform.