A collection of pre-trained medical image models in PyTorch. This repository aims to provide a unified and easy-to-use interface for comparing and deploying these models.

• STU-Net (STU-Net-S, STU-Net-B, STU-Net-L, STU-Net-H) pre-trained on TotalSegmentator, CT-ORG, FeTA21, BraTS21 (more datasets are WIP).
• SAM-Med3D (SAM-Med3D) pre-trained on SA-Med3D-140K.
• Other pre-trained medical image models are WIP. (You can request support for your model in Issues.)

You can use this cmd to install this toolkit via pip:

pip install medim

For developers, you can install in the editable mode via:

git clone https://github.com/uni-medical/MedIM.git
cd MedIM
pip install -e .

First, let us import medim.

import medim

You have four ways to create a PyTorch-compatible model with create_model:

1. use models without pretraining

model = medim.create_model("STU-Net-S") 

2. use local checkpoint

model = medim.create_model(
            "STU-Net-S",
            pretrained=True,
            checkpoint_path="../tests/data/small_ep4k.model") 

3. use checkpoint pre-trained on validated datasets (will automatically download it from HuggingFace)

model = medim.create_model("STU-Net-B", dataset="BraTS21")

4. use HuggingFace url (will automatically download it from HuggingFace)

model = medim.create_model(
            "STU-Net-S",
            pretrained=True,
            checkpoint_path="https://huggingface.co/ziyanhuang/STU-Net/blob/main/small_ep4k.model") 

Tips: you can use MEDIM_CKPT_DIR environment variable to set custom path for medim model downloading from huggingface.

Then, you can use it as you like.

input_tensor = torch.randn(1, 1, 128, 128, 128)
output_tensor = model(input_tensor)
print("Output tensor shape:", output_tensor.shape)

More examples are in examples.

• support more pre-training of STU-Net on different datasets.
• support more pre-trained medical image models.
• An easy-to-use interface compatible with MONAI/nnU-Net is still under development. Once developed, you will be able to deploy medical image models more elegantly within the Python/PyTorch ecosystem.