/MSCA-Net

code for our MSCA-Net

Primary LanguagePython

MSCA-Net: Multi-Scale Contextual Attention Network for Skin Lesion Segmentation

Data preparation

We cropped the ISIC 2018 dataset to 224*320 and saved it in npy format, which can be downloaded from Baidu web disk.

link: https://pan.baidu.com/s/1bIVUdzYG_7tuwalbI4Y8Ww

password: c36c

Place the downloaded npy files in the "data" directory and unzip them. The decompression format is as follows:

/data/ISIC2018_npy_all_224_320/image/

​		ISIC_0000000.npy

​		ISIC_0000001.npy

​		...

​		ISIC_0016072.npy

/data/ISIC2018_npy_all_224_320/label/

​		ISIC_0000000_segmentation.npy

​		ISIC_0000001_segmentation.npy

​		......

​		ISIC_0016072_segmentation.npy

Train and Test

Our program is easy to train and test, just need to run "main_train.py".

python main_train.py

Performance on ISIC 2018

Method Para(M) Flops (G) JI DSC ACC
FCN 15.31 21.98 78.66±0.41 86.80±0.32 95.04±0.32
U-Net 34.53 71.61 81.69±0.50 88.81±0.40 95.68±0.29
U-Net++ 36.63 151.59 81.87±0.47 88.93±0.38 95.68±0.33
AttU-Net 34.88 72.81 81.99±0.59 89.03±0.42 95.77±0.26
DeepLabv3+ 39.76 47.34 82.32±0.35 89.26±0.23 95.87±0.23
DenseASPP 35.37 42.63 82.53±0.55 89.35±0.37 95.89±0.28
BCDU-Net 28.8 171.50 80.84±0.57 88.33±0.48 95.48±0.40
Focus-Alpha 26.36 41.92 81.92±0.63 88.93±0.41 95.84±0.44
DO-Net 24.75 122.45 82.61±0.51 89.48±0.37 95.78±0.36
CE-Net 29.00 9.75 82.82±0.45 89.59±0.35 95.97±0.30
CPF-Net 30.65 8.78 82.92±0.52 89.63±0.42 96.02±0.34
MSCA-Net (Ours) 27.09 12.88 84.18±0.38 90.52±0.26 96.41±0.29

Reference

If you find our work is helpful for your research, please consider to cite:

@article{sun2023msca,
  title={MSCA-Net: Multi-scale contextual attention network for skin lesion segmentation},
  author={Sun, Yongheng and Dai, Duwei and Zhang, Qianni and Wang, Yaqi and Xu, Songhua and Lian, Chunfeng},
  journal={Pattern Recognition},
  pages={109524},
  year={2023},
  publisher={Elsevier}
}