Tencent/MedicalNet

MedicalNet – 3‑D Medical Image Transfer Learning

What it is

• A PyTorch implementation of the Med3D paper ("Transfer Learning for 3D Medical Image Analysis").
• Provides a collection of large‑scale 3‑D medical imaging datasets (23 public datasets) and a suite of pre‑trained 3‑D ResNet models (depths 10‑200).
• Includes scripts to fine‑tune these models on new tasks such as lung segmentation, nodule classification, or brain‑MRI segmentation.

Why it matters

• Training 3‑D CNNs from scratch requires huge annotated volumes, which are scarce in medicine. MedicalNet’s pre‑trained weights give a strong starting point, dramatically improving convergence speed and accuracy (e.g., Dice ↑ 27 % on lung segmentation).

Key components

• models/resnet.py – definition of 3‑D ResNet variants.
• datasets/ – data‑preprocessing utilities (e.g., MRBrainS18 loader).
• train.py / test.py – ready‑to‑run training and evaluation loops supporting multi‑GPU.
• pretrain/ – downloadable .pth files for each architecture.
• utils/logger.py – simple logging.

Typical workflow

1. Install Python 3.7, PyTorch 0.4.1, CUDA 9.0, cuDNN 7.0.5.
2. pip install -r requirements.txt.
3. Download the data and pre‑trained weights (Google Drive / Tencent Weiyun) and unzip into the repository.
4. Run python train.py --gpu_id 0 (or 0 1 for multi‑GPU) to fine‑tune on your dataset.
5. Evaluate with python test.py --gpu_id 0 --resume_path <model>.pth.tar --img_list data/val.txt.

Performance snapshot

Model	Params (M)	Inference time (s) on Tesla P40
3D‑ResNet10	14.36	0.18
3D‑ResNet18	32.99	0.19
3D‑ResNet34	63.31	0.22
3D‑ResNet50	46.21	0.21
3D‑ResNet101	85.31	0.29
3D‑ResNet152	117.51	0.34
3D‑ResNet200	126.74	0.45

Fine‑tuned results (LungSeg Dice, NoduleCls accuracy) show consistent gains over training from scratch across all depths.

License

• MIT License (see LICENSE).

Citation

@article{chen2019med3d,
  title={Med3D: Transfer Learning for 3D Medical Image Analysis},
  author={Chen, Sihong and Ma, Kai and Zheng, Yefeng},
  journal={arXiv preprint arXiv:1904.00625},
  year={2019}
}

Roadmap / TODO

• Completed: pre‑trained 3‑D ResNets, transfer‑learning code, multi‑GPU support.
• Planned: 3‑D efficient architectures (MobileNet, ShuffleNet) and 2‑D medical pre‑trained models.

Who might use this

• Researchers needing a strong baseline for any 3‑D medical imaging task (CT, MRI, PET).
• Clinicians building prototype segmentation or classification tools.
• Developers integrating 3‑D CNNs into larger medical AI pipelines.

---

MedicalNet is a genuine, research‑grade toolbox for 3‑D medical image deep learning, not a generic AI‑related list or tutorial.