GLIMS: Attention-guided lightweight multi-scale hybrid network for volumetric semantic segmentation

This repository contains the code of GLIMS.

GLIMS ranked in the top 5 among 65 unique submissions during the validation phase of the Adult Glioblastoma Segmentation challenge of BraTS 2023.

Installation

Clone the repository

git clone https://github.com/yaziciz/GLIMS.git
cd GLIMS

Install the required dependencies

With your virtual environment activated, install the project's dependencies:

pip install -r requirements.txt

Usage Instructions

Running the Main Script

The GLIMS model can be trained by the given script on the BraTS 2023 dataset:

python main.py --output_dir <output_directory> --data_dir <data_directory> --json_list <json_list_file> --fold <fold_id>

Validation

By using the pre-trained model, the validation phase can be performed as follows:

python post_validation.py --output_dir <output_directory> --data_dir <data_directory> --json_list <json_list_file> --fold <fold_number> --pretrained_dir <pretrained_model_directory>

Testing with Model Ensembles

To test GLIMS by using the ensemble method on the unannotated BraTS 2023 dataset, the following script can be used:

python test_BraTS.py --data_dir <validation_data_directory> --model_ensemble_1 <model_1_path> --model_ensemble_2 <model_2_path> --output_dir <output_directory>

The model_ensemble_1 and model_ensemble_2 variables represent the fold 2 and fold 4 models, as indicated in our challenge submission paper on arXiv.

Citations

GLIMS: Attention-guided lightweight multi-scale hybrid network for volumetric semantic segmentation
Image and Vision Computing, May 2024
Journal Paper, arXiv

@article{yazici2024glims,
  title={GLIMS: Attention-guided lightweight multi-scale hybrid network for volumetric semantic segmentation},
  author={Yazici, Ziya Ata and Oksuz, Ilkay and Ekenel, Hazim Kemal},
  journal={Image and Vision Computing},
  pages={105055},
  year={2024},
  publisher={Elsevier},
  doi={https://doi.org/10.1016/j.imavis.2024.105055}
}

Attention-Enhanced Hybrid Feature Aggregation Network for 3D Brain Tumor Segmentation
Accepted to the 9th Brain Lesion (BrainLes) Workshop @ MICCAI 2023
arXiv

@article{yazici2024attention,
  title={Attention-Enhanced Hybrid Feature Aggregation Network for 3D Brain Tumor Segmentation},
  author={Yazici, Ziya Ata and Oksuz, Ilkay and Ekenel, Hazim Kemal},
  journal={arXiv preprint arXiv:2403.09942},
  year={2024}
}

Thank you for your interest in our work!

We are also deeply grateful to the MONAI Consortium for their MONAI framework, which was instrumental in the development of GLIMS.