Disentanglement of content and style features in multi-center cytology images via contrastive self-supervised learning

Multi-center cervical cytology images have various image styles due to the differences in staining and imaging techniques, which pose a significant challenge to the performance of automated cervical cancer diagnosis tools. We propose a dual-head network architecture that explicitly disentangles image features into content and style features, and applies contrastive self-supervised learning to a large number of unlabeled images, achieving enhanced generalization across various styles. We pretrain our model on 1,024,855 images cropped from 3,561 whole slide images (WSIs), and visualize the features using t-distributed stochastic neighbor embedding (t-SNE) method, demonstrating the effectiveness of our method in distinguishing between content and style features. In the downstream task, we evaluate our model on 192,123 binary-classified images with 10 styles, and achieve the best accuracy among all methods for every style. Across the 10 different data sources, our method attained an average accuracy of 80.4%, outperforming all other comparative methods by 3% to 17%, demonstrating our method's potential to enhance the performance and robustness of automated cytology image analysis in multi-center settings.