Yet another STAin NORmalization Method: Point Set Registration for Color Space Alignment in Histological Images

Advancements in automatic histology image analysis over the past two decades have been driven by machine learning and deep learning technologies. However, the generalization capability of these models is challenged by domain shift. This phenomena arises when models are applied to images with visual characteristics different than those used in the model training. Preprocessing methods, such as stain normalization techniques, have been used previously in histological image analysis pipelines and could be used to mitigate domain shift. Despite decades of research, the impact of these methods remains a topic of ongoing discussion.In this work we present YSTANORM (Yet another STain NOrmalization Method): a color normalization algorithm that aligns pixel data points in color or optical density space. The method leverages point set registration to compute a transformation between the color distributions of a reference and target images. Evaluation of YSTANORM is carried out by assessing the performance of state-of-the-art segmentation models specialized in the histological domain. These models were tested on unnormalized images, images normalized with several stain normalization methods, and images processed with YSTANORM. Our results indicate that most of the stain normalization methods could negatively impact model performance. Notably, YSTANORM can improve the performance of the segmentation models, selecting an appropriate reference image.