willer-lu / DCSAU-Net

ICIP 2022 Submission

DCSAU-Net: A Deeper and More Compact Split-Attention U-Net for Medical Image Segmentation

By Qing Xu, Wenting Duan and Na He

Requirements

1. pytorch==1.10.0
2. pytorch-lightning==1.1.0
3. albumentations==0.3.2
4. seaborn
5. sklearn

Dataset

To apply the model on a custom dataset, the data tree should be constructed as:

    ├── data
          ├── images
                ├── image_1.png
                ├── image_2.png
                ├── image_n.png
          ├── masks
                ├── image_1.png
                ├── image_2.png
                ├── image_n.png

CSV generation

python data_split_csv.py --dataset your/data/path --size 0.9 

Train

python train.py --dataset your/data/path --csvfile your/csv/path --loss dice --batch 16 --lr 0.001 --epoch 150 

Evaluation

python eval_binary.py --dataset your/data/path --csvfile your/csv/path --model save_models/epoch_last.pth --debug True

Acknowledgement

The codes are modified from ResNeSt, U-Net