A weakly supervised multiple instance learning approach for classification of Breast cancer HER-2 status using whole slide images
Zheng Gao, Tian Dong Chen, Bei Yang · 2024
Breast cancer, as the most common malignant tumor in women, poses great harm to women's health. Among breast cancer patients, there is a relatively high proportion of HER2-positive cases, making the prediction of HER2 status extremely important. Currently, many deep learning methods based on Whole Slide Images (WSIs) have been proposed, but these methods still face the problem of imbalanced datasets, with significant differences in the number of samples in each class, causing the model to overfit categories with an advantage in quantity while ignoring minority categories. Therefore, to address the above issues, this paper proposes a weakly supervised Dual-tier Cluster-based Multiple Instance Learning (DCMIL) framework for the classification of breast cancer images. This framework uses ResNet-101 as the backbone and optimizes the pseudo bag partitioning process by introducing clustering mechanisms in the DTFD algorithm. A series of experiments were conducted on the HER2 dataset. The results showed that our model outperforms the latest algorithms in terms of performance.