Stain Transformation of Immunohistochemistry Images Using DINO

Stain transformation refers to the process of converting an already stained tissue image into another type of stain using deep learning techniques. This paper addresses the problem of transforming H&E- stained tissue images into IHC-stained images for breast cancer examination. The challenge in this problem lies in the difficulty of creating datasets where the corresponding image pairs are perfectly aligned at the pixel level, as it is impossible to chemically re-stain an already stained image. Therefore, research in this field focuses on techniques that overcome the inconsistencies between the source and target images. In this paper, we apply DINO, a ViT-based knowledge distillation network that has proven effective in various applications, to the pyramid-pix2pix network for this problem. The DINO net- work extracts invariant feature vectors against partially transformed images, and our experiments demonstrate that this property is useful in addressing image transformation tasks where pixel-level alignment is compromised.