Breast cancer full-slice image classification method combining self-supervision and weak supervision learning

The invention discloses a multi-instance breast cancer full-slice pathological image classification method based on combination of a self-supervision method and a weak supervision method. The method is divided into two stages of MoBY self-supervision contrast learning and weak supervision multi-instance learning based on Transform. In the first stage, a comparative learning strategy is used for training on a large amount of label-free data. Swin Transform is selected as a backbone model, through a self-attention mechanism and hierarchical feature representation, changes and differences of different breast cancer pathology images are better adapted, and tissue features of the breast cancer pathology images are preliminarily learned. And in the second stage, the weight initialization model in the first stage is utilized, each full-slice image is regarded as a packet by adopting a multi-example learning method, and the generated small slices are regarded as examples in the packet. A Top-n example with the maximu