microsoft/siim-workshop-2026

microsoft/siim-workshop-2026

Language: Jupyter Notebook

License: MIT

Stars: 5

Forks: 0

Open issues: 4

Created: 2026-06-09T16:13:40Z

Pushed: 2026-06-16T04:50:29Z

Default branch: main

Fork: no

Archived: no

README:

How to Actually Use Foundation Models: From Classification to Agents Learning Lab

SIIM 2026 Learning Lab. This repository supports a hands-on lab on building radiology AI agents in VS Code with GitHub Copilot. The lab is organized into three chapters, each in its own folder under [chapters/](chapters/).

Budget about 30 min setup (do this before arriving) + 60 min content across the three chapters.

What you'll build

1. Chapter 1 — Simple agent + skill ([chapters/chapter1-simple-agent/](chapters/chapter1-simple-agent/)): use a general-purpose LLM with a prompt-based skill to draft structured radiology reports. 2. Chapter 2 — Using foundation models with agents ([chapters/chapter2-foundation-models/](chapters/chapter2-foundation-models/)): wrap MedImageInsight (zero-shot + a trained adapter) in an MCP server the agent can call. 3. Chapter 3 — Patient context ([chapters/chapter3-knowledge-sources/](chapters/chapter3-knowledge-sources/)): ground agent outputs in a (synthetic) patient history.

Workshop layout

• [chapters/](chapters/) — one folder per chapter. Each has a README.md, a starter/ skeleton, and an answer-key/ reference.
• [samples/](samples/) — 30 sample chest X-rays (NIH 2017) you can hand to the agent.
• [eval-results/](eval-results/) — reference scoring numbers we ship for the part-to-part comparisons.

> The data and patient records in this repo are synthetic or de-identified and are not for clinical use. The images are from the 2017 NIH Kaggle dataset. Ground-truth report text in eval-results/ is model-generated from NIH labels for reference scoring only; it is not radiologist-written.

Prerequisites

Install and verify the following before the workshop. Expect about 30 minutes of setup. Do this before arriving — Wi-Fi and time will be tight on the day.

• VS Code (latest stable): https://code.visualstudio.com/
• GitHub Copilot and GitHub Copilot Chat extensions, signed in with an account that has Copilot access.
• Git, and a local clone of this repository opened as your VS Code workspace.
• uv (https://docs.astral.sh/uv/) — Python package manager. Install user-scope (no admin required):
• macOS / Linux: curl -LsSf https://astral.sh/uv/install.sh | sh
• Windows (PowerShell): powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Then run uv sync from the repo root.

MedImageInsight credentials

Chapter 2 hits a shared MedImageInsight endpoint. Credentials (MI2_MODEL_ENDPOINT and MI2_MODEL_API_KEY) are distributed at the start of the session. You'll set them up in Chapter 2 Step 1; the working path will be mcp/medimageinsight-server/.env at workspace root.

Quick check

The workshop/ directory at repo root holds shared Python utilities for the workshop, including a small script to verify your env is healthy.

uv --version # 0.x.x
uv sync # one-time, ~1–2 min on a good connection
uv run workshop/check_env.py # expected: "Workshop env OK — Python 3.11"

If any command above fails, resolve it before arriving.