by Min Liu, Yubin Han, Jiazheng Wang, Can Wang*, Yaonan Wang, and Erik Meijering.
- This is the Pytorch implementation for our paper 'LSKANet: Long Strip Kernel Attention Network for Robotic Surgical Scene Segmentation', accepted by IEEE Transactions on Medical Imaging (TMI), in 2023.12.
- We proposed a surgical scene segmentation network named Long Strip Kernel Attention network (LSKANet), which includes two newly designed modules, Dual-block Large Kernel Attention module (DLKA) and Multiscale Affinity Feature Fusion module (MAFF). Besides, the hybrid loss with Boundary Guided Head (BGH) is proposed to help the network segment indistinguishable boundaries effectively. Our LSKANet achieves new state-of-the-art results on three datasets with different surgical scenes (Endovis2018, CaDIS, and MILS) with relative improvements of 2.6%, 1.5%, and 3.4% mIoU, respectively.
- We evaluate the proposed LSKANet on three datasets:Endovis2018, CaDIS, and a self-built dataset called Minimally Invasive Laparoscopic Surgery dataset (MILS), which will be released in the future.
We provide some visualization resluts here, more resluts can be found in paper.
- Visualization results on Endovis2018 (12 classes)
- Visualization results on CaDIS (Task Ⅲ, 25 classes)
- Visualization results on MILS (8 classes)
We used these packages/versions in the development of this project.
* PyTorch 1.10.0
* torchvision 0.12.0
* mmcv 1.6.1
* mmsegmentation 0.24.1
* opencv-python 4.5.3
Before training, please download the dataset you need and rename them following mmseg/datasets/endovis2018.py
and mmseg/datasets/cadis.py
.
- Switch folder
cd ./tools/
- Use
python train.py
to start the training - Parameter setting and training script refer to
/work_dirs/0_LSKANet/LSKANet_XXXX.py
- Use
python test.py
to start the inferencing - Visualization results can be found in
/tools/test_out/
We build our code on MMsegmentation. Thanks original authors for their impressive work!
For further question about the code or paper, please contact Yubin Han:hyb_hnu@163.com.