semantic segmentation of Medical Images

Posted on

发展流程

  • edge detection filters and mathematical methods
  • machine learning approaches extracting hand-crafted features
  • deep learning techniques

network structure

1
- ## CNN

«««< HEAD

CNN

  • patch-wise prediction

Fully Convolutional Network（FCN）

======= ## CNN

  • patch-wise prediction

## Fully Convolutional Network（FCN）

44040e01e9f8f65a10c32c5a25935419d0a7b321

  • pixel-wise prediction from the full-sized image
  • up-sampling the output feature map
  • 保护纹理信息，混合粗粒度层和细粒度层
  • Cascaded FCN（以肝部问题为例，先分割肝，然后在肝上分割病灶）

Encoder-decoder Network

  1. U-Net
    • a contracting path to capture context
    • a symmetric expanding path that enables precise localization
    • skip connections
    • weighted BCE

  2. DCAN

  3. V-Net
    • Residual connections
    • dice loss
  4. DenseNet-based

Attention-based

Adversarial-based

Recurrent Neural Network-based

Transfer Learning

loss function

challenges and limitations