Skip to main content
. 2023 Apr 4;18(4):e0283562. doi: 10.1371/journal.pone.0283562

Fig 4. Intelligent sampling.

Fig 4

All the patches from the tumor region of a WSI are scored by the pretrained CNN-scorer and then rearranged according to their DSs in descending order. The top K patches are sampled as the most discriminative for this WSI. The sampled patches are fed into the same CNN as a feature extractor (i.e., without the classification head) to be embedded into feature vectors. The top K extracted feature vectors act as a bag of instances for subsequent training of an MIL model.