MediScan AI — Medical Triage Agent on AMD MI300X

MediScan AI is a production-grade multi-agent medical triage system built entirely on AMD Instinct MI300X GPU using the ROCm 6.0 open-source stack. The system accepts medical imaging inputs — chest X-rays, MRI, and CT scans — and routes them through three specialized AI agents that operate in a coordinated pipeline to deliver structured clinical triage reports in under 500 milliseconds. Agent 01 — Intake & Preprocessing Agent: Handles DICOM parsing, image normalization to 512×512 resolution, and scan type classification. Runs entirely on AMD Developer Cloud with zero proprietary dependencies. Agent 02 — Pathology Detection Agent: Runs BiomedCLIP, a biomedical vision-language model from Microsoft and UCSF, via PyTorch on ROCm. Detects findings such as Pleural Effusion, Pneumothorax, Pneumonia, Cardiomegaly, and Pulmonary Infiltrates with confidence scoring. Benchmarked at 423ms inference on AMD MI300X — datacenter-grade speed on open hardware. Agent 03 — Clinical Report Generator: Uses LLaMA 3 to convert raw model findings into structured, human-readable clinical triage reports with risk classification (High / Medium / Low) and recommended clinical actions. Business Value: Radiology departments globally face critical bottlenecks — radiologist shortages, scan backlogs, and delayed triage decisions. MediScan AI addresses this directly by providing instant AI-assisted pre-screening, allowing clinicians to prioritize critical cases faster. The system is deployable on AMD Developer Cloud infrastructure, making it cost-accessible for Technical Differentiation: This project demonstrates that AMD ROCm + MI300X is fully production-capable for real medical AI workloads — matching inference speed and model accuracy achievable on proprietary GPU stacks, while remaining fully open-source. GPU utilization ran at 94% sustained compute, 19.2GB VRAM utilized, at 396W power draw — all monitored live within the application dashboard.