VLFAT

January 20, 2026 ยท View on GitHub

This repository contains the source code of the following paper that is accepted for MICCAI 2023(Available at arXiv): Transformer-based end-to-end classification of variable-length volumetric data, Marzieh Oghbaie, Teresa Araujo, Taha Emre, Ursula Schmidt-Erfurth, Hrvoje Bogunovic

The proposed network deploys Transformers for volume classification that is able to handle variable volume resolutions both at development and inference time.

Proposed Approach for 3D volume Classification

Alt text

The main models are available at model_zoo/feature_extrc/models.py. weights are now available at release: https://github.com/marziehoghbaie/VLFAT/releases/tag/weights

Installation

Please check INSTALL.md for installation instructions.

Training

For OLIVES dataset, the list of samples should be provided in a .csv file under dataset to annotation_path_test field. The file should at least includes sample_path,FileSetId,label,label_int,n_frames. On Duke dataset, however, give the directory of the samples arranged like the following to the dataloader is sufficient: subset/class.

python main/Smain.py --config_path config/YML_files/VLFAT.yaml

Evaluation

  • Simple Test with confusion matrix: set the train: false and allow_size_mismatch: false under train_config in the corresponding config file.
python main/Smain.py --config_path config/YML_files/FAT.yaml
  • Calculate AUC of all the models. A list of the corresponding configs should be provided:
python main/Stest_AUC.py
  • Robustness Analysis: The test is run for num_test under model_inputs in the corresponding config file and the results are saved in the corresponding log file. This test only works for
python main/Stest_robustness.py --config_path config/YML_files/VLFAT.yaml

Note that the list of possible volume resolutions can be changed in the main/Stest_robustness.py file (e.g. volume_resolutions = [5, 15, 25, 35, 45]).

Acknowledgement

This repository is built using the timm library, Pytorch and Meta Research repositories.

License

This project is released under the MIT license. Please see the LICENSE file for more information.

Citation