RadYOLO is a 3D YOLO11-based deep learning model for object detection, classification, and instance segmentation in medical images. It is designed to process MRI or CT volumes efficiently on GPU and CPU-only systems. In our paper published at the 2nd MICCAI Workshop on Efficient Medical AI, we could show that RadYOLOs detection performance is on par with nnU-Net for lesion detection and coarse organ detection. At the same time it is 8-46x faster on a GPU and over 100x faster on a CPU, solving most detection tasks within a couple seconds.
RadYOLO was developed by the Fraunhofer Institute for Digital Medicine MEVIS.
If you use RadYOLO, please cite our paper:
@inproceedings{geissler2026radyolo,
title={RadYOLO: Computationally Efficient 3D Object Detection and Segmentation in CT and MRI},
author={Geissler, Kai and Müller-Groh, Laurens and Meine, Hans},
booktitle={<enter when published>},
year={2026},
}- Install dependencies from repository root:
uv build .or
pip install .- Prepare a dataset in RadYOLO format (images, labels, dataset.yaml).
- Create data splits.
- Optionally analyze/preprocess data and propose a model config.
- Start training:
uv run -m radyolo train <path-to-config-yaml>Detailed guides are available in documentation/:
- Dataset preparation: dataset-preparation.md
- Data splits: data-splits.md
- Analyze and preprocess data (optional): analyze-and-preprocess.md
- Propose model configuration (optional): propose-model-configuration.md
- Training quickstart: training.md
- Inference and evaluation: inference-and-evaluation.md
- Build docker image: docker.md
The software is not qualified for use as a medical product or as part thereof. Provided 'as is' without specific verification or validation.
RadYOLO is licensed under the AGPL-3.0 license.
Kai Geissler, kai.geissler<at>mevis.fraunhofer.de, Fraunhofer Institute for Digital Medicine MEVIS, Germany