StairwayToStain: A Gradual Stain Translation Approach for Glomeruli Segmentation

Ali Alhaj Abdo, Islem Mhiri, Zeeshan Nisar, Barbara Seeliger, Thomas Lampert

Keywords: Digital Pathology, Stain Translation, Domain Adaptation, Domain Shift

Abstract: Unsupervised domain adaptation (UDA) has advanced digital pathology by enabling knowledge transfer across clinical contexts. Although promising, most UDA frameworks focus on adapting source-labeled data to target unlabeled data directly in a one-off way. However, translating stains from information-poor domains to information-rich ones can lead to a domain shift problem due to the large discrepancy between domains. To address this issue, we propose StairwayToStain (STS), an unsupervised gradual stain translation framework that uses intermediate stains to bridge the gap between the source and target stain. Our method is grounded in three main phases: (i) measuring the domain shift between different stains, (ii) defining a translation path, and (iii) performing the gradual stain translation. Our method demonstrates its efficacy in improving glomeruli segmentation when translating from immunohistochemical (IHC) to histochemical stains, as well as between different IHC stains. Comprehensive experiments on stain translation using STS demonstrate competitive results in comparison with its variants and state-of-the-art direct UDA methods. Moreover, we are able to generate additional stains during the translation process. Our method presents the first framework for gradual domain adaptation in stain translation.

Submission Number: 17