Fig 6. The proposed BCR-Net model.

All patches from the WSI’s tumor region (annotated in green) are fed into a pretrained CNN-scorer. The patches are rearranged according to their DSs as previously defined (see Eq 1) in descending order. The top K patches are sampled and embedded by the same feature extractor inherited from the CNN-scorer (Fig 4). The output K feature vectors are treated as a bag of instances and aggregated through attention-based pooling. As per the attention-based pooling, the attention weights a1, a2, …, aK are produced by the ANN. And then, a weighted sum is conducted to aggregate the feature vectors with their attention weights. As the final representation of the WSI, the output meta-instance is classified by a fully connected layer (FCN) and a probability score will indicate the final pre-diction for the WSI.