A multiple instance learning Framework with Feature Fine-Tuning for Whole Slide image classification

The upgrading of computer hardware and the development of artificial intelligence are driving the advancement of computational pathology. However, Whole Slide images (WSIs) which contain billions of pixels, typically only have labels at the image level. As a result, multiple instance learning (MIL) is commonly used for pathology image analysis. Most existing methods utilize attention mechanisms to aggregate instance-level features. However, relying solely on attention mechanisms to aggregate instance-level features can constrain the model to local optima and risk overlooking key regions or features. To address this issue, we draw inspiration from the principles of Generative Adversarial Networks (GANs) and introduce adversarial networks into the multi-instance learning framework. By incorporating adversarial networks, we introduce a degree of intervention into the attention mechanism when aggregating bag-level features, allowing for the fine-tuning of these features. We conducted experiments on two publicly available datasets, TCGA-NSCLC and TCGA-RCC, and demonstrated the effectiveness and feasibility of our approach.