🧠 KonfAI

KonfAI is a modular deep learning framework for medical imaging built around YAML-driven workflows.

It lets you define complete pipelines, from data loading to prediction and evaluation, through configuration instead of orchestration scripts.

• reproducible workflows with explicit configs and outputs
• modular components for datasets, transforms, models, losses, and metrics
• designed for research, experimentation, and agent-driven workflows

KonfAI has been used in several top-performing challenge projects: 🔗 SynthRAD2025 – Task 1 • 🔗 SynthRAD2025 – Task 2 • 🔗 CURVAS PDACVI 2025 • 🔗 TrackRAD 2025 • 🔗 Panther • 🔗 CURVAS

Paper:

KonfAI: A Modular and Fully Configurable Framework for Deep Learning in Medical Imaging

---

🚀 Why KonfAI?

Most frameworks focus on models. KonfAI focuses on pipelines.

• 🧩 Compose full workflows from modular components
• 🔁 Iterate without rewriting Python scripts
• 📦 Turn experiments into reusable KonfAI Apps
• 🤖 Use KonfAI as a backend for LLM-driven experimentation through KonfAI-MCP

---

⚡ Quickstart

Install and run your first workflow:

git clone https://github.com/vboussot/KonfAI.git
cd KonfAI
python -m pip install -e .
cd examples/Segmentation

python -m pip install -U "huggingface_hub[cli]"
hf download VBoussot/konfai-demo \
  --repo-type dataset \
  --include "Segmentation/**" \
  --local-dir Dataset
mv Dataset/Segmentation/* Dataset/
rmdir Dataset/Segmentation
rm -rf Dataset/.cache

This downloads a small public demo subset and prepares the layout expected by the example:

Dataset/
├── 1PC006/
│   ├── CT.mha
│   └── SEG.mha
└── ...

• CT is the input image
• SEG is the segmentation label map

Then launch the first training run:

konfai TRAIN -y --gpu 0 --config Config.yml

Then:

konfai PREDICTION -y --gpu 0 --config Prediction.yml --models Checkpoints/SEG_BASELINE/<checkpoint>.pt
konfai EVALUATION -y --config Evaluation.yml

If you do not have a GPU available, replace --gpu 0 with --cpu 1.

Edit these files first:

• Config.yml → training
• Prediction.yml → exported outputs
• Evaluation.yml → metrics on saved predictions

Notebook entry points:

• examples/Segmentation/Segmentation_demo.ipynb
• examples/Synthesis/Synthesis_demo.ipynb

For editable installs and optional extras such as server or cluster, see:

• docs/source/getting-started/installation.md

---

🤖 Agent-Ready by Design

KonfAI is designed to serve as a deterministic backend for LLM-driven experimentation.

Through KonfAI-MCP Server, agents can:

• inspect datasets
• generate or refine configurations
• launch experiments
• analyze results and iterate

All executions remain:

• reproducible
• structured
• grounded in YAML workflows

👉 KonfAI bridges the gap between LLM reasoning and real experimental execution.

---

📦 KonfAI Apps

A KonfAI App is a self-contained workflow package built with KonfAI.

It can expose:

• inference
• evaluation
• uncertainty estimation
• full pipelines
• fine-tuning

Apps live in apps/ and can be used through:

Interface	Entry point
🖥️ CLI	konfai-apps
🐍 Python API	konfai_apps.KonfAIApp
🌐 Remote server	konfai-apps-server + konfai-apps --host ...
🧠 3D Slicer	SlicerKonfAI

Use Apps when a workflow is already stable and you want a cleaner user-facing interface than the low-level YAML CLI.

---

📚 Documentation

The README is only the entry point. The full documentation is available here:

• https://konfai.readthedocs.io/en/latest/

---

🐳 Docker

KonfAI ships a Docker setup for CLI-oriented workflows.

• Dockerfile: docker/Dockerfile
• guide: docs/source/usage/docker.md
• image: vboussot/konfai

Example:

docker run --rm -it \
  --gpus all \
  -v "$(pwd):/workspace" \
  -w /workspace \
  vboussot/konfai TRAIN --gpu 0 -c examples/Synthesis/Config.yml

---

🤝 Contributing

Contributions are welcome.

Typical ways to help:

• improve examples and notebooks
• clarify documentation
• add tests for real user paths
• extend models, transforms, or apps

Local setup:

git clone https://github.com/vboussot/KonfAI.git
cd KonfAI
python -m pip install -e .
python -m pip install pytest pre-commit

Useful commands:

pytest -q
pre-commit run --all-files
make -C docs html

Contributor guide:

• docs/source/contributing.md