Fig. 2: StainLUT framework for training the self-supervised stain normalization model. | npj Digital Medicine

Fig. 2: StainLUT framework for training the self-supervised stain normalization model.

From: Self-supervised stain normalization empowers privacy-preserving and model generalization in digital pathology

Fig. 2

The input images are first preprocessed in two steps: color randomization (CR) and grayscaling. Then they are input to the Predictor and Encoder to obtain the parameter of the perceptual image scene and the context map (CM) of the perceptual image content, respectively. The former is used for weighted fusion of basis stainLUTs to obtain context-aware stainLUT, and the latter is cascaded with the corresponding three-channel grayscale images to obtain four-channel grayscale-context concatenated images. Finally, the four-channel images are fed into the context-aware stainLUT to obtain the output images by quadrilinear interpolation operations. The to-be-output and input images are trained under the guidance of the loss function to obtain the final StainLUT model for inference. More details regarding this scheme execution are provided in the Methods section.

Back to article page