Algorithm 1:Spatial attention mechanism
Inputs:The input map of the two branches features of PET/CT and CTχil,i=1,2
Output:SAl1:
1:h1l=concat(χ1l,χ2l)/* Add the feature maps of PET/CT and CT */
2:hmean=AvgPool(h1l)/*avg-pooling*/
3:hmax=MaxPool(h1l)/*max-pooling*/
4:f=concat(hmean,hmax)/* Concatenate the feature map of avg-pooling and max-pooling */
5:β=Conv3×3(f)/*3×3 convolution operation */
6:z=σ(β)/* After sigmoid,the feature map becomes C×H×1*/
7:SAl=z×h1l+h1l /* The feature map of sigmoid is multiplied with the original feature and then add*/
End
Algorithm 2:Channel attention mechanism
Inputs:The input features map of the three branches of PET/CT,CT and PETχil,i=1,2,3
Output:CAl(F)
1:χhybridl=χ2l⊕χ3l /*Add the feature maps of PET and CT,χhybirdl*/
2:χl=χhybridl+χ3l /*Concatenate the feature map of χhybridl and PET/CT*/
3:hl,gl=AvgPool(χl),MaxPool(χl)/* Avg-pooling and max-pooling on the feature maps respectively */
4:α=σ(MLP(hl)+MLP(gl))/*Perform MLP operations on hl,gl separately*/
5:CAl(F)=α×χl+χl
Set citation alerts for the article
Please enter your email address
CancelConfirm