Utilizing Mask-Guided Cross-Image Attention for Zero-Shot In-Silico Histopathologic Image Generation with a Diffusion Model

Utilizing generative models to create insilica data presents an economical alternative to the traditional methods of staining, imaging, and annotating images in the computational pathology workflow. Specifically, appearance transfer diffusion models enable the generation of images without the need for model training, making the process swift and efficient. While originally developed for natural images, these models can transfer foreground objects from a source to a target domain, with less emphasis on the background. In computational pathology, however, every aspect of an image, including the background, can be crucial for understanding the tumor micro-environment. In this study, we adapted an appearance transfer diffusion model to align with the demands of computational pathology by adjusting the AdaIN feature statistics in the denoising process. The effectiveness of this modified method was demonstrated through its application to epithelium segmentation. The results demonstrated superior performance compared to the baseline approach, indicating that the number of manual annotations necessary for model training could be reduced by 75% without sacrificing accuracy. The authors expect that this research will promote the use of zero-shot diffusion models in computational pathology.