Segformer3D is a light-weight and efficient hierarchical Transformer designed for 3D volumetric segmentation. It calculates attention across multiscale volumetric features, and avoids complex decoders. Instead it uses a simple yet effective all-MLP decoder to aggregate local and global attention features to produce highly accurate segmentation masks.
The implementation of SegFormer3D architecture is in architectures/segformer3d.py. The experimentatl setup and implementaiton details are throuhgly explained in our paper.
Step by step guideline for BraTs is provided below. Other datasets are coming soon.
- (April 15, 2024): SegFormer3D code & weights are released for BRaTs.
Follow steps 1-3 to run our pipeline.
Make sure you have conda installed and run the following lines of commands in your terminal:
git clone https://github.com/OSUPCVLab/SegFormer3D.git
cd SegFormer3D
conda create -n "segformer3d" python=3.11.7 ipython==8.12.0
conda activate segformer3d
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=11.8 -c pytorch -c nvidia
# make sure <which python> and <which pip> points to the created environment directory
python -m pip install -r requirements.txtFor convenience, we call SegFormer3D repo's home directory SEGFORMER3D_HOME.
First we need to make two folders (I call them train, BraTS2017_Training_Data) using the terminal:
cd SEGFORMER3D_HOME
mkdir "./data/brats2017_seg/brats2017_raw_data/train"
mkdir "./data/brats2017_seg/BraTS2017_Training_Data"Download Brats2017. Open the zip file and extract the content into the train directory of the repo. Once the files are moved properly, you should see the following folder structure:
SEGFORMER3D_HOME/data/brats2017_seg/brats2017_raw_data/
│
├───train
│ ├──imageTr
│ │ └──BRATS_001_0000.nii.gz
│ │ └──BRATS_001_0001.nii.gz
│ │ └──BRATS_001_0002.nii.gz
│ │ └──BRATS_001_0003.nii.gz
│ │ └──BRATS_002_0000.nii.gz
│ │ └──...
│ ├──labelsTr
│ │ └──BRATS_001.nii.gz
│ │ └──BRATS_002.nii.gz
│ │ └──...
│ ├──imageTs
│ │ └──BRATS_485_000.nii.gz
│ │ └──BRATS_485_001.nii.gz
│ │ └──BRATS_485_002.nii.gz
│ │ └──BRATS_485_003.nii.gz
│ │ └──BRATS_486_000.nii.gz
│ │ └──...
The train folder contains 484 MRI scans of the raw BraTs2017 data.
For the sake of accelerating the training process, we perform all the base preprocessing (zero cropping, collating the modalities, etc) prior to actually training the model. Run
brats2017_seg_preprocess.py code which saves the preprocessed volumes into the BraTS2017_Training_Data folder. Having the data preprocess will save a lot of time during the training.
Simply run (this will take a while):
cd SEGFORMER3D_HOME
# assuming conda environment is still activated
python ./data/brats2017_seg/brats2017_raw_data/brats2017_seg_preprocess.py
# you need around 22 GB to store the preprocessed data After this step, the dataset is saved in BraTS2017_Training_Data and is ready to be used for training.
Note: BraTs2021 dataloader and preprocessor are also available in our repo. You can use them the same way as explained for BraTs2017.
Inside SEGFORMER3D_HOME/data/brats2017_seg you see two csv files called train.csv and validation.csv which define the splits for training and validation data for your experiment. For generating different splits you can go to SEGFORMER3D_HOME/data/brats2017_seg/brats2017_raw_data/datameta_generator.
You can also find kfold splits inside SEGFORMER3D_HOME/data/brats2017_seg/brats2017_raw_data/datameta_generator. For kfold experiment you have to change the fold_id inside the train_dataset_args and val_dataset_args section of your experiment's config.yaml. Use same fold id for both train_dataset_args and val_dataset_args. The default value is fold_id: null which flags the dataloader to read train.csv and validation.csv. You can change it to your desired fold id. For example, setting fold_id: 1 in train_dataset_args and val_dataset_args will make the dataloader to read validation_fold_1.csv and train_fold_1.csv during the training. Just make sure your desired splits are present in SEGFORMER3D_HOME/data/brats2017_seg.
In order to run an experiment, we provide a template folder placed under SEGFORMER3D_HOME/experiments/brats_2017/template_experiment that you can use to setup your experiment. Simply copy the the template folder (maybe rename it to something more meaningful) and run your experiment on a single GPU with:
cd SEGFORMER3D_HOME
cd ./experiments/brats_2017/your_experiment_folder_name/
# the default gpu device is set to cuda:0 (you can change it)
accelerate launch --config_file ./gpu_accelerate.yaml run_experiment.py
# you can modify the cuda device id by changing the value of gpu_ids attribute in the gpu_accelerate.yaml
# for example gpu_ids: '1' will run the experiment on cuda:1You might want to change the hyperparameters (batch size, learning rate, weight decay etc.) of your experiment. For that you need to edit the config.yaml file inside your experiment folder.
As the experiment is running, the logs (train loss, vlaidation loss and dice score) will be written to the terminal. You can log your experiment on wandb
(you need to setup an account there) if you set mode: "online" in the wandb_parameters section of the config.yaml. The default value is mode: "offline". If you want to log the result to your wandb account, put your wandb info into the wandb_parameters section of the config.yaml and your entire experiment will be logged under your wandb entity (e.g. pcvlab) page.
During the experiment, the model checkpoints (optimizer, model_parameter etc.) will be saved in SEGFORMER3D_HOME/experiments/brats_2017/your_experiment_folder_name/model_checkpoints/best_dice_checkpoint/ upon each improvement of the metric (Dice Score). Model parameters is saved as pytorch_model.bin.
You can download Segformer3d model weight trained on BraTs2017 from [GoogleDrive].
You can download SegFormer3D visualization results from this link. This is a standalone folder that you might want to tinker with. viz_meta.yaml contains the hex color code for each dataset and other meta information for viz.py to visualize the data. By chaning the random seed number in viz.py you will get different frames visualized from the 3d volume input. We did not set any specific policy behind how we choose the frames to visualize because the volumetric inputs are considerably different with each other. Alternatively, you can modify viz.py according to your preference.
We got the visualization for nnFormer and UNETR form nnFormer visualization results.
We benchmark SegFormer3D both qualitatively and quantitatively against the current state-of-the-art models such as nnFormer, UNETR and several other models on three widely used medical datasets: Synapse, BRaTs2017, and ACDC. The quatitative results are based on reported average Dice Score.
BraTs is a collection of Magnetic Resonance Imaging (MRI) scans that contains 484 MRI images with four modalities, FLAIR, T1w, T1gd, and T2w. Data were collected from 19 institutions with ground-truth labels for three types of tumor subregions: edema (ED), enhancing tumor (ET) and nonenhancing tumor (NET).
The Synapse dataset provides 30 annotated CT images with thirteen abdominal organs.
ACDC is a dataset of 100 patients used for 3D volumetric segmentation of the left (LV) and right (RV) cardiac ventricles and the myocardium (Myo).
If you liked our paper, please consider citing it
@inproceedings{perera2024segformer3d,
title={SegFormer3D: an Efficient Transformer for 3D Medical Image Segmentation},
author={Perera, Shehan and Navard, Pouyan and Yilmaz, Alper},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={4981--4988},
year={2024}
}@article{perera2024segformer3d,
title={SegFormer3D: an Efficient Transformer for 3D Medical Image Segmentation},
author={Perera, Shehan and Navard, Pouyan and Yilmaz, Alper},
journal={arXiv preprint arXiv:2404.10156},
year={2024}
}
