EyeGuard: A Multimodal AI Early Screening System for Ophthalmic Diseases Targeting South Asian Populations

EyeGuard is a mobile-first multimodal AI platform designed specifically for the South Asian region. By integrating retinal image analysis and clinical data, it enables early detection of five major ophthalmic diseases: diabetic retinopathy, glaucoma, macular degeneration, and cataracts. The system combines explainable AI technology with teleophthalmology services, aiming to address the screening challenges of ophthalmic diseases caused by uneven distribution of medical resources in South Asia, allowing patients to access convenient screening and diagnosis support via smartphones.

Globally, ophthalmic diseases are the main causes of visual impairment and blindness, but early intervention can effectively control most cases. However, in regions with scarce medical resources like South Asia, the distribution of professional ophthalmologists is uneven. Traditional screening relies on professional equipment and experienced doctors, making large-scale promotion difficult. The popularity of smartphones and maturity of AI technology provide a new direction to solve this problem—capturing fundus images via mobile devices, combining AI screening, and connecting to telemedicine.

EyeGuard is a multimodal AI screening system designed for low-resource environments, adopting a four-layer architecture:

1. Mobile Application Layer: Developed based on Flutter/Dart, supporting image collection, metadata entry, offline TFLite inference, data upload, and result display, covering Android and iOS cross-platform.
2. API Gateway Layer: Built with FastAPI/Django REST, responsible for request orchestration, JWT/RBAC authentication, and routing.
3. AI Inference Layer: The core pipeline includes image encoder, metadata encoder, cross-attention fusion, classifier, and GradCAM visualization.
4. Data Storage Layer: PostgreSQL stores user profiles and screening records; cloud storage (AWS S3/Google Cloud) saves encrypted images, integrating services like Stripe payment and PMDC doctor verification.

The highlight of EyeGuard's AI pipeline lies in cross-attention fusion:

• Image Encoder: Uses Vision Transformer (ViT-B/16) or Swin Transformer to extract deep features from fundus images.
• Metadata Encoder: Encodes clinical data such as age, blood pressure, and blood glucose into feature vectors via fully connected layers.
• Cross-Attention Fusion: Learns complex correlations between image features and clinical data (e.g., the association between retinopathy and high blood sugar).
• Classification and Interpretability: Outputs probability distributions for five diseases; GradCAM generates heatmaps to highlight key decision-making areas. Supported disease types: Diabetic retinopathy, glaucoma, macular degeneration, cataracts, normal.

• Offline Inference: Implements local screening via TensorFlow Lite, adapting to areas with poor network connectivity and protecting privacy (sensitive data processed locally).
• Telemedicine: High-risk cases trigger push/SMS/email notifications and generate PDF reports; users can book video consultations with PMDC-verified doctors, integrating Stripe payment.
• Data Security: HTTPS transmission, AES-256 encrypted storage, JWT/RBAC access control, complying with medical data regulations.

EyeGuard addresses the problem of heavy ophthalmic disease burden but scarce doctors in South Asia:

• Decentralized Screening: Patients do not need to travel long distances for medical treatment; primary health workers can use AI-assisted screening.
• Efficiency Improvement: Pre-screens high-risk cases to optimize doctors' diagnosis and treatment processes.
• Reduction of Preventable Vision Loss: Helps groups like diabetic patients detect lesions in a timely manner.

As an undergraduate graduation project, EyeGuard needs further improvement:

• Clinical Validation: Requires large-scale real-world testing of system performance.
• Regulatory Approval: Needs approval from medical regulatory authorities for official deployment.
• Localization: Add multilingual support to adapt to different regions.
• Expand Disease Coverage: Can support detection of more ophthalmic diseases in the future. This project demonstrates the application potential of AI technology in regions with scarce medical resources, committed to making high-quality medical services accessible to more people.