MSNet&M2SNet

MSNet: Automatic Polyp Segmentation via Multi-scale Subtraction Network

Xiaoqi Zhao, Lihe Zhang, Huchuan Lu
⭐ arXiv »

M2SNet: Multi-scale in Multi-scale Subtraction Network for Medical Image Segmentation

Xiaoqi Zhao, Hongpeng Jia, Youwei Pang, Long Lv, Feng Tian, Lihe Zhang, Weibing Sun, Huchuan Lu
⭐ arXiv »

• Official repository of "Automatic Polyp Segmentation via Multi-scale Subtraction Network" MICCAI-2021.
• Official repository of "M2SNet: Multi-scale in Multi-scale Subtraction Network for Medical Image Segmentation".
• :boom: We won the second place (2/100) in the MICCAI 2022 Challenge: Glaucoma Oct Analysis and Layer Segmentation (GOALS).

Datasets

• Image Polyp Segmentation: (training dataset) Google Drive / BaiduYunPan(342v); (testing dataset) Google Drive / BaiduYunPan(4e10)
• Video Polyp Segmentation: (training dataset) Google Drive / BaiduYunPan(4b45); (testing dataset) Google Drive / BaiduYunPan(ki4a)
• COVID-19 Lung Infection: (training dataset) Google Drive / BaiduYunPan(uoiy); (testing dataset) Google Drive / BaiduYunPan(r867)
• Breast Ultrasound Segmentation: Google Drive / BaiduYunPan(409p)

Results

• MSNet: Google Drive / BaiduYunPan(j3i8)
• M2SNet: Google Drive / BaiduYunPan(a60m)

Trained Model

• You can download the trained MSNet model at Google Drive / BaiduYunPan(ps6e).
• You can download the trained M2SNet model at Google Drive / BaiduYunPan(5vo3).
• You can download Res2Net weights at Google Drive / BaiduYunPan(w46l)

Highlight

Novel Segmentation Architectures

Efficient Intra-Layer Multi-scale Subtraction Design

Efficient Inter-Layer Multi-scale Subtraction Structure

Training-free Loss Network

Low FLOPs (comparisons under the Res2Net-50 backbone)

Prerequisites

• Python 3.7
• Pytorch 1.8.1
• OpenCV 4.6
• Numpy 1.19
• Apex

Training/Inference/Testing

• set the cfg in train.py:

    Dataset.Config(datapath='', savepath='', mode='train', batch=16, lr=0.05, momen=0.9, decay=5e-4, epoch='')
    %the number of training epochs settings in the polyp segmentation, COVID-19 Lung Infection, breast tumor segmentation and OCT layer segmentation are 50, 200, 100         and 100, respectively.
    python train.py

• Run prediction_rgb.py (can generate the predicted maps)
• Run test_score.py (support 10 binary segmentation evaluation metrics: MAE, maxF, avgF, wfm, Sm, Em, M_dice, M_iou, Ber, Acc)

TODO LIST

• [ ] 3D verison MSNet training.

• [ ] Support different backbones (VGGNet, MobileNet, ResNet, Swin, etc.).

• [ ] Diverse Medical Image Segmentation

  • [x] Polyp
  • [x] COVID-19 Lung Infection
  • [x] Breast tumor
  • [x] OCT Layer
  • [ ] Prostate
  • [ ] Cell Nuclei
  • [ ] Liver
  • [ ] Retinal Vessel
  • [ ] Skin Lesion
  • [ ] Lung
  • [ ] Pancreas
  • [ ] Hippocampus
  • [ ] Heart
  • [ ] BrainTumour

BibTex

@inproceedings{MSNet,
  title={Automatic polyp segmentation via multi-scale subtraction network},
  author={Zhao, Xiaoqi and Zhang, Lihe and Lu, Huchuan},
  booktitle={MICCAI},
  pages={120--130},
  year={2021},
  organization={Springer}
}

@article{M2SNet,
  title={M $\^{}$\{$2$\}$ $ SNet: Multi-scale in Multi-scale Subtraction Network for Medical Image Segmentation},
  author={Zhao, Xiaoqi and Jia, Hongpeng and Pang, Youwei and Lv, Long and Tian, Feng and Zhang, Lihe and Sun, Weibing and Lu, Huchuan},
  journal={arXiv preprint arXiv:2303.10894},
  year={2023}
}