Algorithm 2 FRDCNN model

Input:  two training sets: S1, S2;

test set: T;

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

transferred weights from the published study: Wu

foreground masks: MS2;

cell counts label: Cconcatenate(S1, S2).

Procedure:

1.    L3 = f3(A (W3, S2), MS2))

train a set of weights W3 which was initialized randomly to minimize the loss function L3.

2.   L4 = f4 (V (W4, A (W3, concatenate (S1, S2)), Cconcatenate(S1, S2)))

W3 was initialized from Step 1 and fixed to untrainable.

W4 was trained to minimize the loss function L4.

Output:  predicted cell counts: cDRDCNN.

cFRDCNN = V (W2, U (W1, T)