
Autonomous Oncology Board (AOB) is a multi-agent AI system that simulates a real multidisciplinary tumour board—pathologist, clinical researcher, and oncologist—then produces a structured Patient Management Plan (TNM staging, biomarker gating, NCCN categories, and citations). AOB starts from histopathology patches plus basic case metadata. The Pathologist agent uses Prov‑GigaPath (ViT) to embed and classify tissue, highlight suspicious regions with explainability overlays, and report calibrated confidence (with uncertainty when needed). The Researcher agent performs retrieval‑augmented generation (RAG) over a pre‑indexed local corpus (e.g., NCCN guidance + TCGA/PubMed papers), returning an Evidence Bundle where each recommendation is tied to sources. The Oncologist agent synthesizes both outputs into a final plan and runs a debate loop: the Researcher critiques missing tests or guideline constraints (e.g., biomarker status), and the Oncologist revises—showing a visible revision diff and a consensus score. This architecture is enabled by AMD Instinct MI300X (192GB unified VRAM), allowing multiple large models plus KV cache to remain resident without swapping. The demo includes a live VRAM dashboard and an H100 “OOM” comparison bar to make the memory math tangible. Disclaimer: research prototype, not for clinical use.
10 May 2026