MediVision: Dermatology & Wound Care AI Assistant

Access to quality dermatology care is limited in many regions, leaving patients without timely diagnosis for skin conditions and wounds. MediVision addresses this gap by providing an AI-powered assistant that analyzes skin and wound images combined with patient symptom descriptions to deliver structured, actionable insights. Built for the AMD Developer Hackathon 2026 (Track 3: Vision & Multimodal AI), MediVision uses the Qwen2.5-VL-7B vision-language model deployed via vLLM on AMD Instinct MI300X GPUs through the AMD Developer Cloud. The model performs multimodal analysis, processing both images and text to identify conditions such as abrasions, dermatitis, eczema, ringworm, psoriasis, burns, and cellulitis. Key Features: - Multimodal Analysis: Combines skin/wound images with patient symptom descriptions for comprehensive assessment. - Multilingual Support: Outputs in 6 languages — English, Tiếng Việt, 中文, Español, Français, and 日本語 — making it accessible globally. - Structured Output: Provides diagnosis, severity level (Low/Medium/High/Urgent), treatment recommendations, and confidence scores. - AMD Hardware Powered: Leverages AMD MI300X + ROCm for fast, efficient inference. - Gradio Frontend: Hosted on Hugging Face Spaces, providing an easy-to-use interface with minimal dependencies. Tech Stack: - Vision Model: Qwen/Qwen2.5-VL-7B-Instruct - Inference: vLLM (OpenAI API compatible) - Hardware: AMD Instinct MI300X + ROCm - Host: AMD Developer Cloud - Frontend: Gradio 5.29 on HF Spaces Live Demo: https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/medivision-ai-agent Source Code: https://github.com/HiImSunny/medivision-ai-agent