SAStainDiff: Self-supervised stain normalization by stain augmentation using denoising diffusion probabilistic models

With the development of computer-aided detection/diagnosis, histopathological images become increasingly important for cancer diagnosis and prognosis. However, different stain styles in histopathological images arise from the difference in stain techniques, operator skills, and scanner specifications. These stain styles reduce the robustness of computer-aided detection/diagnosis algorithms. Existing stain normalization methods often suffer from poor generalization ability and the issue of information loss. In this paper, we propose a new self-supervised diffusion probabilistic modeling approach for stain normalization with stain augmentation training strategy and rescheduled sampling strategy, termed SAStainDiff. Specifically, we employ stain augmentation to simulate different stain styles and learn any stain distribution through diffusion models in a self-supervised manner while preserving the histopathological structure. We employ rescheduled sampling strategy that selects fewer sampling step sizes and a different initial sampling point. This reduces the inference time, which is comparable to mainstream methods, while keeping the performance. We conduct experiments on mutual stain normalization between breast cancer images scanned by two different scanners. Additionally, we explore the application of stain normalization in lymphoma classification and colon gland segmentation. Experimental results demonstrate that our method exhibits excellent generalization capabilities and adapts well to different tissue textures and stain styles without retraining, achieving satisfactory performance in terms of both speed and quality. Our proposed SAStainDiff method can improve the accuracy of disease diagnosis and subsequent analysis, ultimately benefiting clinical practice and advancing medical research. The code and sample data are publicly available on GitHub.