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

TL;DR

This paper tackles the problem of classifying breast cancer histopathology images under limited annotations by introducing Dynamic Cross-Scale Swin Transformer (DCS-ST). The approach combines a Dynamic Window Predictor, Cross-Scale Attention Module, and a diffusion-based semi-supervised learning pipeline with pseudo-labeling to enhance feature representation across scales while leveraging unlabeled data. The authors provide extensive experiments on BreakHis, Mini-DDSM, and ICIAR2018 BACH, showing improvements in AUC-ROC, balanced accuracy, F1, and Cohen's Kappa compared to strong baselines, especially at low labeling rates. The work advances medical image analysis by enabling robust, multi-scale histopathology classification with minimal annotation effort, offering potential clinical utility for early breast cancer detection.

Abstract

Deep learning methods have shown promise in classifying breast cancer histopathology images, but their performance often declines with limited annotated data, a critical challenge in medical imaging due to the high cost and expertise required for annotations.

Figures (7)

• Figure 1: Flowchart of the Dynamic Cross-Scale Swin Transformer (DCS-ST) pipeline for breast cancer histopathology image classification with limited annotations. Input images are split into 5% labeled and 95% unlabeled data, processed via a Dynamic Window Predictor and pseudo-labeling strategy, followed by backbone feature extraction and classification. This pipeline effectively leverages both labeled and unlabeled data for robust performance.
• Figure 2: Images of Breast Cancer Histopathology from the BreakHis Dataset at Four Magnifications
• Figure 3: Images of Breast Cancer Histopathology from the ICIAR2018 Dataset
• Figure 4: Images of Breast Cancer Histopathology from the Mini-DDSM Dataset
• Figure 5: Comparison of the Confusion Matrices of the BreakHis Dataset