Joint Dense Residual and Recurrent Attention Network for DCE-MRI Breast Tumor Segmentation

Figure 5

Comparison of the standard U-Net and the proposed recurrent attention U-Net architecture. (a) The standard U-Net consists of a left-right symmetrical encoder and decoder. In the encoder, feature maps are continuously downsampled by max-pooling to extract high-level semantic information. In the decoder, the feature map resolution is gradually recovered by transposed convolution. The rich low-level semantic information of the encoder is concatenated through a skip structure, which compensates for the information loss caused by the downsampling and upsampling process. (b) The proposed recurrent attention U-Net. A dense residual module was added at the end of each stage (except stage 1) of the U-Net encoder, and the output of the module was used for skip connections. Furthermore, our proposed recurrent attention module replaced the convolution blocks at each stage of the U-Net decoder. Note that each rectangle in the picture represents a feature matrix. The number above each rectangle represents the number of feature matrix channels and the number on the lower left represents the resolution of the feature map.

(a)

(b)