Figure 1.

Overall structure of SECP-Net. After multiple convolutions and down-sampling, there are different sizes and levels of information captured in different stages for input images. We utilize SEC and pyramidal networks to fuse the information and extract more contributing features to segmentation by the channel attention mechanism of the SE block [18]. Then the information flow is transmitted to the decoder through a skip-connection. After the extraction of the primary part, we combine original inputs and the probability distribution from the primary network according to the auto-context, which is sent to the secondary U-Net for further segmentation and more accurate results.