🏥 MediAgent
Autonomous Multi-Agent Medical Imaging Analysis — AMD Instinct MI300X
AMD Developer Hackathon 2026 · Vision & Multimodal AI Track Built by Ramyar — Sulaymaniyah, Iraq
What It Does
MediAgent runs a 5-agent AI pipeline that analyzes medical images (X-ray, MRI, CT, DICOM) and generates formal radiology reports with differential diagnoses, ICD-10 codes, and FHIR R4 export — entirely on AMD hardware.
No cloud APIs. No OpenAI. No Nvidia. Pure AMD MI300X + ROCm + vLLM.
⚠️ Demo Mode
This Space runs in demo mode — the full pipeline UI works and all 5 agents animate live, but no real inference is performed since the AMD Instinct MI300X backend is not available on HuggingFace's free hardware.
See the video demo for live inference on real AMD hardware.
Live inference requires: AMD Instinct MI300X · ROCm · vLLM · Qwen multimodal
The 5-Agent Pipeline
| Agent | Role |
|---|---|
| INTAKE | Validates input, normalizes clinical language, safety triage |
| VISION | Multimodal image analysis via Qwen on AMD MI300X |
| RESEARCH | KB cross-reference, differential diagnoses, ICD-10 codes |
| REPORT | ACR/NICE format radiology report synthesis |
| CRITIC | QA peer-review, quality scoring, disclaimer enforcement |
INTAKE + VISION run in parallel to minimize latency.
Key Features
- Real-time SSE streaming pipeline with per-agent timers
- DICOM (.dcm) file support with metadata extraction
- 15-condition medical knowledge base with ICD-10 mapping
- FHIR R4 DiagnosticReport export
- Client-side PDF export
- Post-report clinical Q&A (ClinicalAdvisorAgent)
- Live AMD GPU metrics (util, VRAM, temp, power)
- Hard-enforced clinical safety rules
Tech Stack
- GPU: AMD Instinct MI300X
- GPU Software: ROCm
- Inference: vLLM (ROCm build) + Qwen multimodal
- Backend: FastAPI + Uvicorn
- Frontend: Vanilla JS + Tailwind CSS + Chart.js + SSE
GitHub
Full source code, architecture docs, and README: https://github.com/Ramyar2007/mediagent
This system is a decision support tool. All outputs must be reviewed by a licensed radiologist before any clinical decisions are made.
Model tree for lablab-ai-amd-developer-hackathon/MediAgent
Base model
Qwen/Qwen3.6-27B