Algorithm 3 CRDCNN model

Input:  two training sets: S1, S2;

test set: T;

models: U-net (U), self-built autoencoder (A), VGG (V);

dot density maps: DS1;

foreground masks: MS2

cell counts label: Cconcatenate(S1, S2)

Procedure:

1.   L5 = f5 (V (W5, concatenate (U (W1, concatenate (S1, S2)), A (W3, concatenate (S1, S2))), Cconcatenate(S1, S2)))

W1 was transferred from the DRDCNN model and fixed to untrainable.

W3 was transferred from the FRDCNN model and fixed to untrainable.

W5 was trained to minimize the loss function L5.

Output:  predicted cell counts: cDRDCNN.

cERDCNN = V (W5, concatenate (U (W1, T), A (W3, T)))