图像分割论文:Efficient, high-performance pancreatic segmentation using multi-scale feature extraction
Auther: Moritz Knolle, Institution:Technical University of Munich, Publish year: 2020
论文地址:https://arxiv.org/abs/:2009.00872
代码地址:https://github.com/TUM-AIMED/MoNet
Key point:
- 在医学分割领域,提高模型参数利用率
- 通过多尺度图像特征,实现高效胰腺分割
- U-Net架构以及重复扩张卷积(RDDC)
- 在参数较少(减少了95%的参数量)且推理速度较快的情况下,在集外测试数据上性能优于U-Net,att-U-Net
Network:
Fig1 MoNet
Fig2 RDDC block
2D U-Net based结构,输入为256*256,而RDDC,可实现多尺度特征提取
Performance:
Fig3 Comparision with U-Net and att U-Net