This is the offical implementation of our paper: Attention2Minority: A salient instance inference-based multiple instance learning for classifying small lesions in breast cancer whole slide images. Paper

Camelyon16 dataset, torch, torchvision, tensorboard, openslide, PIL, pandas, numpy, scikit-learn, tqdm, opencv

Extract the coordinates of the top-left corner of each patch from CAM16 raw slides:

$ python extraction.py --slidedir <>

Or use your own:

data
    ├── pts
          ├── cam16l1p224s224
                            ├── slide_1.npy
                            ├── slide_2.npy
                            └── ...

Encoding patches from CAM16 raw slides using Resnet50(pretrained on ImageNet, and truncated at the third block):

$ python encoding_pts.py --slidedir <>

Or use your own:

data
   ├── feats
           ├── cam16res
                      ├── train
                              ├── normal
                                       ├── slide_1.npy
                                       ├── slide_2.npy
                                       └── ...
                              └── tumor
                                      └── ...
                      └── test
                             ├── normal
                                      └── ...
                             └── tumor
                                     └── ...

Learn representative negative instances (i.e., Key set)

$ python keyset_lrn.py -t 100

Or download the learned key set.

$ python sii.py -k 150

$ python train_cv.py -r 0.3 --keys sm_sort.npy --code cam16res_siimil --data cam16_sii

Download the pretrained models.

Sii selected instances can also boost performance of other MIL models. It currently works better on non-contextual models.

$ python eval_cv.py -r 0.3 --keys sm_sort.npy --code cam16res_siimil --data cam16_sii