Abstract
The complexity in digital pathology image analysis arises from the inherent variability in histopathological slides including tissue specimen differences and stain variations. While publicly available datasets are increasing they primarily focus on hematoxylin and eosin (H&E) staining yet pathologists often require analysis across multiple stains for comprehensive diagnosis. The poor cross-domain generalization of deep learning models intensifies these challenges making it unfeasible to implement without exhaustive annotations in each stain. This manual curation process is time-consuming and requires the creation of sufficiently large and precise datasets. In this work we address these challenges focusing on breast cancer across four crucial stains: ER PR HER2 and Ki-67. Given the necessity of cell-level information for diagnosis we concentrate on cell detection tasks. Leveraging unsupervised domain adaptation techniques we bridge the gap between publicly available annotated H&E datasets and unlabeled data in other stains. Utilizing detection transformers we demonstrate the superiority of adversarial feature learning over source-only and image-level generative methods. Our work contributes to improving digital pathology analysis by enabling robust and efficient pipelines across multiple stains. By leveraging unsupervised domain adaptation we facilitate the development of accurate computer-aided diagnosis pipelines capable of handling stain variations and enhancing diagnostic accuracy in practical settings.