Deploying image-based screening on mobile devices requires balancing predictive performance, model size, and inference latency. This paper presents IrisHealth, a mobile deployment for on-device lung-sector iris screening based on deep learning. The system performs binary classification on a harmonized lung-sector iris corpus under a subject-disjoint protocol and runs interactive on-device inference entirely on the phone. Experimental results show that ConvNeXt-Tiny achieves the highest accuracy of 96.21% and an F1-score of 94.18%, and is retained as an offline accuracy reference. For on-device deployment, MobileNetV3-Small provides a practical trade-off, achieving 94.14% accuracy with 1.07 M parameters and an end-to-end latency of 389.7 ± 16.5 ms (∼2.6 FPS). IrisHealth is developed as an engineering research prototype and is not intended for clinically validated medical diagnosis.

Iris image analysis;Disease classification;Deep learning;On-device AI