Diffusion Models for Generative Histopathology

Niranjan Sridhar, Michael Elad, Carson McNeil, Ehud Rivlin, Daniel Freedman

Abstract: Conventional histopathology requires chemical staining to make tissue samples usable by pathologists for diagnosis. This introduces cost and variability and does not conserve the tissue for advanced molecular analysis of the sample. We demonstrate the use of conditional denoising diffusion models applied to non-destructive autofluorescence images of tissue samples in order to generate virtually stained images. To demonstrate the power of this technique, we would like to measure the perceptual quality of the generated images; however, standard measures like the Frechet Inception Distance (FID) are inappropriate for this task, as they have been trained on natural images. We therefore introduce a new perceptual measure, the Frechet StainNet Distance (FSD), and show that our model attains significantly higher FSD than competing pix2pix models. Finally, we also present a method of quantifying uncertain regions of the image using the variations produced by diffusion models.