DCS-ST for Classification of Breast Cancer Histopathology Images with Limited Annotations

To address this, we propose the Dynamic Cross-Scale Swin Transformer (DCS-ST), a robust approach designed for classifying breast cancer histopathology images under constrained annotation settings. DCS-ST processes input images by leveraging a small set of labeled data alongside a larger pool of unlabeled data, employing a pseudo-labeling strategy to generate high-confidence labels for training. A dynamic window predictor adaptively adjusts attention window sizes across scales, enhancing the Swin Transformer backbone's ability to capture both local and global features.