Unsupervised single-domain generalization for tissue classification via progressive domain transformation.

Tissue classification is one of the fundamental tasks in computational pathology, but domain shifts in digital pathology images limit the generalization of classification models. Domain generalization has emerged as a leading solution to address this gap, with related research often using multiple public datasets to demonstrate model generalization ability across different sources. To further explore this, we introduce the GDPH-CRC-HE-MS dataset, consisting of 101 H&E-stained colorectal cancer slides from Guangdong Provincial People's Hospital, scanned by 1 to 6 different scanners. In this study, we propose an unsupervised single-domain progressive generalization (USD-PG) framework, which incorporates two progressive data transformations: style progressive data transformation (Style-PDT) and spatial progressive data transformation (Spatial-PDT). This approach prevents unreasonable texture and color changes caused by completely random transformations during the early training stages. We evaluate the generalization ability of the USD-PG framework on the new GDPH-CRC-HE-MS dataset as well as the publicly available NCT-CRC-HE-100K dataset. Our results demonstrate that USD-PG achieves superior performance in single-source domain generalization for tissue classification, effectively handling both scanner-based and data-source domain shifts. It highlights the potential of USD-PG for enhancing domain generalization in tissue classification and its applicability in clinical settings. The source code and the released datasets are available at: https://github.com/linjiatai/USD-PG.