DrRetina — AI Diabetic Retinopathy Agent

DrRetina is an end-to-end AI diagnostic system for Diabetic Retinopathy (DR) detection, built for ophthalmologists and medical professionals. DR affects over 537 million diabetics worldwide and is the leading cause of preventable blindness, yet most clinics in South Asia and Africa lack access to specialist screening tools. Our system solves this through three integrated layers: VISION ENGINE: A fine-tuned ViT-MAE (facebook/vit-mae-base) model trained on the APTOS 2019 dataset using AMD Instinct MI300X GPUs via ROCm. The model achieves Cohen's Kappa of 0.9097 — surpassing the WHO DR screening benchmark of 0.80 and our own target of 0.85. GradCAM heatmaps highlight the exact retinal lesions driving each diagnosis, building clinical trust. AGENTIC LAYER: A LangChain ReAct agent powered by Qwen3-8B with 5 specialized clinical tools. The agent generates structured diagnostic reports and answers follow-up clinical questions with full diagnosis context. Reports are automatically generated in 6 languages — English, Urdu, Arabic, Hindi, Spanish, and French — with RTL text support for Urdu and Arabic via WeasyPrint and Google Noto Fonts. AMD INFRASTRUCTURE: The entire fine-tuning pipeline runs on AMD Instinct MI300X via ROCm 7.2 and PyTorch. Training completes in approximately 5.3 minutes per 50 epochs. The system is deployed as a Hugging Face Space with a FastAPI inference microservice backend. DrRetina is a three-track submission covering Vision & Multimodal AI, Fine-Tuning on AMD GPUs, and AI Agents & Agentic Workflows — making it one of the most complete medical AI systems in this hackathon with verified, quantified accuracy metrics.