The unofficial implementation of UNETR: Transformers for 3D Medical Image Segmentation on Pytorch
Output of my implementation. (A) Ground Truth; (B) Prediction
- Locality of the receptive fields in convolutional layers still limits their learning capabilities to relatively small regions. Combining self-attention modules with convolutional layers has been proposed to improve the non-local modeling capability.[1]
- UNETR utilizes a contracting-expanding pattern consisting of a stack of transformers as the encoder which is connected to a decoder via skip connections. As commonly used in NLP, the transformers operate on 1D sequence of input embeddings.[1]
UNETR Architecture Figure from Official Paper
- Python 3.6+
pip install -r requirements.txt
- BRATS 2016 and 2017[2] datasets were used for training.
- Dataset were downloaded and splitted training/validation automatically by MONAI[3] from Medical Decathlon challenge[4].
- Training process can be started with following command.
python main.py --mode train --model_path ./path/to/model.pth --dataset_path ./path/to/dataset
- Dataset will be downloaded to given dataset path if not exist.
- Trained model will be evaluated after the training process.
- Pretrained model(s) can be evaluated with following command too.
python main.py --mode evaluate --model_path ./path/to/model.pth --dataset_path ./path/to/dataset
- After model is trained, inference can be run for single data with following command.
python main.py --mode inference --model_path ./path/to/model.pth --data_path ./path/to/image
- Results will be saved to
./results/<data_name>/for each class (TC/WT/ET).