Fig. 2: The training framework of our DS neural network model.

a The backbone of the DS network G is built on the CUT framework, which combines contrastive learning and adversarial learning. The input is a 2D OCT-SC map X and the output is a digitally stained image G(X) that is compared with a PS image Y from an adjacent slice. b Auxiliary pseudo-supervised learning task. The biophysical module computes the optical density OD(Y) of the PS image Y, which is fed as an input to G. The digitally stained OD image G(OD(Y)) is compared with the original PS image Y during training. c Auxiliary unsupervised cross-modality image registration task. We alternate between optimizing G and a registration network R under different image scales. We fix R while updating G, which provides more informative supervision for R in the next iteration. We use patch-wise losses for training G, and whole slide image (WSI) losses for training R. The red and green channels of the deformation field represent the vertical and horizontal displacement vectors, respectively