Global and local feature reconstruction network-based medical image segmentation method

A global and local feature reconstruction network-based medical image segmentation apparatus, comprising:

• an input module, configured to input a to-be-detected image into a pretrained global and local feature reconstruction network, wherein the global and local feature reconstruction network comprises a feature encoding module, a global feature reconstruction GFR module, and a feature decoding module based on a local feature reconstruction LFR module, and the global feature reconstruction GFR module is embedded in a skip connection between the feature encoding module and the feature decoding module;
• a feature encoding module, configured to perform feature extraction on the to-be-detected image by using the feature encoding module, to obtain feature maps in a plurality of levels;
• a global descriptor generation module, configured to generate a global descriptor for each level of feature map by using the global feature reconstruction GFR module;
• a cross-level descriptor generation module, configured to generate a cross-level global descriptor of the each level of feature map by using the global feature reconstruction GFR module in combination with the global descriptor of the each level of feature map and a global descriptor of a high-level feature map;
• a global feature reconstruction module, configured to predict a reconstruction weight of all pixels in the each level of feature map by using the global feature reconstruction GFR module, multiply the reconstruction weight by the cross-level global descriptor to implement feature reconstruction, and perform addition with the each level of feature map to obtain a global reconstruction feature map of each level;
• and
• a feature decoding module, configured to perform fusion upsampling on an adjacent higher-stage local reconstruction feature map and an adjacent lower-level global reconstruction feature map by using a local feature reconstruction module in the feature decoding module, to obtain a current-stage high-resolution semantic feature map, and upsample a first-stage high-resolution semantic feature map to obtain a target segmentation prediction image.