Official Implementation of "ESFPNet: efficient deep learning architecture for real-time lesion segmentation in autofluorescence bronchoscopic video"
- Install Pytorch:
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
- Install image reading and writting library:
conda install -c conda-forge imageio
- Install image processing library:
pip install scikit-image
- Install library for parsing and emitting YAML:
pip install pyyaml
- Install other packages:
conda install pillow numpy matplotlib
- Install Jupyter-Notebook to run .ipynb file
conda install -c anaconda jupyter
- Download the training and testing dataset from this link: Experiment Dataset
- Extract the folders and copy them under "Endoscope-WL" folder
- The training and testing dataset are ordered as follows in "Endoscope-WL" folder:
|-- TrainDataset
| |-- CVC-ClinicDB
| | |-- images
| | |-- masks
| |-- Kvasir
| |-- images
| |-- masks
|-- TestDataset
| |-- CVC-300
| | |-- images
| | |-- masks
| |-- CVC-ClinicDB
| | |-- images
| | |-- masks
| |-- CVC-ColonDB
| | |-- images
| | |-- masks
| |-- ETIS-LaribPolypDB
| | |-- images
| | |-- masks
| |-- Kvasir
| |-- images
| |-- masks
- The default dataset paths can be changed in "Configure.yaml"
- To randomly split the CVC-ClincDB or Kvasir dataset, set "if_renew = True" in "ESFPNet_Endoscope_Learning_Ability.ipynb"
- To repeat generate the splitting dataset, previous generated folder shold be detelted first
- To reuse the splitting dataset without generating a new dataset, set "if_renew = False"
- Download the pretrained Mixtransformer from this link: Pretrained Model
- Put the pretrained models under "Pretrained" folder
If you think this paper helps, please cite:
@article{chang2022esfpnet,
title={ESFPNet: efficient deep learning architecture for real-time lesion segmentation in autofluorescence bronchoscopic video},
author={Chang, Qi and Ahmad, Danish and Toth, Jennifer and Bascom, Rebecca and Higgins, William E},
journal={arXiv preprint arXiv:2207.07759},
year={2022}
}
Since the training of MixTransformer based network requires a good GPU. One helpful state-of-the-art work compared in this paper without using MixTransformer backbone is CARANet If you also think this work helps, please cite:
@inproceedings{lou2021caranet,
author = {Ange Lou and Shuyue Guan and Hanseok Ko and Murray H. Loew},
title = {{CaraNet: context axial reverse attention network for segmentation of small medical objects}},
volume = {12032},
booktitle = {Medical Imaging 2022: Image Processing},
organization = {International Society for Optics and Photonics},
publisher = {SPIE},
pages = {81 -- 92},
year = {2022},
doi = {10.1117/12.2611802}}