Domain Generalization for Pathological Images Using the Storage Period Information

Yuki Shigeyasu, Shota Harada, Akihiko Yoshizawa, Kazuhiro Terada, Naoki Nakazima, Mariyo Kurata, Hiroyuki Abe, Tetsuo Ushiku, Ryoma Bise

Abstract: This paper brings attention to a new issue in the development of datasets for AI-based pathological image analysis: the domain shift problem stemming from the storage period until digital scanning. Pathological slides that were sliced and stained in the past may also be scanned to prepare a dataset for AI in addition to the newly stained tissue. However, it leads to changes in their appearance and causes domain shift. We propose a domain generalization method that leverages the storage period as sub-domains. Furthermore, we introduce the ordinal adversarial loss that can perform well for ordinal domain classes. The experiments show the effectiveness of using the storage period as a sub-domain and the ordinal adversarial loss.