pooya-mohammadi / unet-skin-cancer

U-Net-based Models for Skin Lesion Segmentation: More Attention and Augmentation

U-Net-based Models for Skin Lesion Segmentation: More Attention and Augmentation

This repository is associated with this paper. In this work, ten models and four augmentation configurations are trained on the ISIC 2016 dataset. Each augmentation configuration is a combination of CutMix, Mosaic, Hair-Augmentation, Hair-Removal, and basic geometrical augmentations.

The performance and overfitting are compared utilizing five metrics. Our results show that the UR50 and the R2U have the highest metrics value, along with two data augmentation scenarios. We also investigate CBAM and AG blocks in the U-Net architecture, which enhances segmentation performance at a meager computational cost. In addition, we propose UPCG, in which pyramid, AG, and CBAM blocks are used in a sequence in the basic U-Net architecture, which significantly surpasses the results of using the two individually and got the better of UR50 and DU models in some situations. The architecture of UPCG is depicted below. We also show that R2UC model can be more successful than its parent, R2U, and other attention-less models in the boundaries and more difficult scenarios.

Our work is hereby available for reproduction purposes. Please follow the succeeding instructions.

Download dataset

To download & preprocess the data, run the following module. In case of URL fetch failure turn on your VPN.

python data/preprocessing.py --train_path data/train --test_path data/test --mask_train_path data/mask_train --mask_test_path data/mask_test

Train the model

python train.py --model <model-name> ...

python train.py --model r2unet --epochs 25 --early_stopping_p 20 --cutmix_p 1 --hair_aug_p 0 --mosaic_p 0 --batch_size 16

Try Jupyter notebook

train-2016.ipynb

Cite as

If you have benefited from our work, please consider citing us with the following Bibtex citation format:

@misc{https://doi.org/10.48550/arxiv.2210.16399,
  doi = {10.48550/ARXIV.2210.16399},
  
  url = {https://arxiv.org/abs/2210.16399},
  
  author = {Kazaj, Pooya Mohammadi and Koosheshi, MohammadHossein and Shahedi, Ali and Sadr, Alireza Vafaei},
   
  keywords = {Image and Video Processing (eess.IV), Computer Vision and Pattern Recognition (cs.CV), Machine Learning (cs.LG), FOS: Electrical engineering, electronic engineering, information engineering, FOS: Electrical engineering, electronic engineering, information engineering, FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {U-Net-based Models for Skin Lesion Segmentation: More Attention and Augmentation},
  
  publisher = {arXiv},
  
  year = {2022},
  
  copyright = {Creative Commons Attribution 4.0 International}
}

unet-skin-cancer's Issues

Import Error on deep_utils

Hey, there!

I am running the train_2016.ipynb file on the google colab and I having import error on every deep_utils functions.

I am assuming that this deep_utils modul is written by you for the projects not the pypi package we are talking about. So, I checked the deep_utils repo of yours but still not find the files and function I need for this project.

How can I get that training notebook working? Any Ideas?