Edge AI Device for Retina Disease Detection

Overview

The system is a portable, self-contained embedded device for point-of-care screening of vision-threatening retinal disease — specifically diabetic retinopathy and cataract. It combines a fundus-imaging optical front end with on-device deep-learning inference so that image capture, preprocessing, model execution, and result presentation all happen locally on the device, without any network or cloud dependency.

Edge AI is central to the design: conventional AI-driven retinal screening pipelines assume workstation-class hardware or cloud inference, which is impractical in low-resource clinical settings. Deploying quantised CNN models on a Raspberry Pi–class platform preserves patient data locality, removes connectivity as a failure mode, and brings unit hardware cost to a level consistent with field-deployable screening.

The demonstrated prototype captures fundus images through a Pi Camera V3 and a 20 D aspheric lens, runs a multi-model DR ensemble and a cataract classifier on-device, and reports a diagnosis with a per-class confidence score on an integrated LCD, driven by a joystick / button interface.

System Architecture

Hardware

3D Prototype Showcase

Software & AI

Application code runs on Raspberry Pi OS in Python 3.10. Image capture and preprocessing are implemented with OpenCV; the user interface is composed with PyGame and rendered directly to the LCD framebuffer. Inference uses TensorFlow Lite for the quantised models, with the training-time toolchain (TensorFlow / Keras, PyTorch for exploratory work) kept off-device.

Five CNN classifiers were trained and deployed. Four architectures target diabetic retinopathy grading (MobileNetV2, EfficientNet-B0, EfficientNet-B3, EfficientNet-B5) against the APTOS 2019 Blindness Detection dataset; a fifth model, based on VGG19, performs binary cataract classification against ODIR-5K.

Trained models are converted to TensorFlow Lite and quantised to reduce memory footprint and per-inference latency without a significant accuracy penalty. At run-time all four DR models execute against the captured frame and their outputs are combined by majority vote, which reduces sensitivity to any single model's misclassification.

Key Engineering Work

• Model optimisation for edge deployment — conversion of five CNN classifiers to TensorFlow Lite with quantisation targeted at the Raspberry Pi 5 CPU inference path
• Multi-model ensemble arbitration — running four DR classifiers per frame and resolving them by majority vote to improve robustness against per-model bias
• Optical / sensor integration — mating a 20 D aspheric lens to the Pi Camera V3 to obtain a usable fundus field of view within a hand-held envelope
• SBC + MCU co-design — offloading analog joystick handling to an Arduino Leonardo so that the SBC's CPU budget is reserved for capture and inference
• Local UI stack — implementing a keyboard/mouse-free interaction model on the LCD with PyGame, suitable for use by non-technical operators
• Enclosure and thermal design — passive-cooled 3D-printed housing that keeps CPU temperatures within a safe steady-state band during continuous inference

Results & Validation

In clinical validation on 13 real patients, the ensemble correctly identified No-DR, Mild, Moderate and Severe cases at average confidence scores above 82%.

Independent ophthalmologist review of the deployment placed overall end-to-end diagnostic accuracy at approximately 70% — consistent with the intended positioning of the device as a low-cost pre-screening tool rather than a definitive diagnostic.

Technical Stack