Fig. 3

Algorithm training. A. ColorAE training. Input image is run through an autoencoder to yield concentration maps of each color (6 distinct mIHC stain colors: yellow, teal, purple, red, black, brown; blue hematoxylin nuclear counterstain; and background.) Two loss functions are applied to ensure that the reconstructed image has the highest fidelity to the original image and expert weak annotations. B. U-Net training. Input image was run through a U-Net. Cross entropy loss function was applied to maximize fidelity to superpixel labels derived from manual annotation of the input image. C. Ensemble method workflow. Input image is run through the autoencoder and U-Net to generate predictions as shown above