Fig. 6

The overall structure of the proposed Fovea-UNet. a. The architecture of Fovea-UNet. Medical input images are first fed into the extracting path and four intermediate features maps are obtained. Then the Fovea Pooling modules take the feature maps as input and yield the output respectively. Lastly, the segmentation mask is acquired by concatenating the output of FP in turn and upsampling layers hierarchically. b. The illustration of Fovea Pooling. The importance-aware module calculates the importance-aware map using the intermediate features as input, and the importance-aware map of each feature provides the basis of the pooling radius. We map the pooling process on the original input images as the illustration, which is shown in the upper part of (b). The closer to warm the color of the picture border is, the more the picture contains detailed information. c. The illustration of the HSIC-Ghost convolution layer. The constraint of the normal convolution layer that generates the distinct intrinsic features is added and then we adopt more cheap operations to ensure the distinction and sufficiency of features