Weakly-Supervised-Segmentation-and-uncertainty-based-Interactive-Correction-on-Whole-Slide-Images

0. Donwload CAMELYON

Download only the tumorous slides of the CAMELYON 16 DATASET. The CAMELYON dataset was orginally created with only 2 folders: train and test. For our purpose we will use 3 folder by splitting the original test folder into a test folder and a validation folder. When the CAMELYON 16 DATASET is downloaded save the annotations and the slides in the directory as described in the config file:

path_slide_tumor_train = os.path.join(PATH_CAMELYON, "train/tumor")
path_slide_tumor_test = os.path.join(PATH_CAMELYON, "test/tumor")
path_annotations_train = os.path.join(PATH_CAMELYON, "annotations")
path_annotations_test = os.path.join(PATH_CAMELYON, "test/annotations")

1. Create Scribbles on each tumorous slide and generate ground truth mask

you can select the percentage of regions you want to scribble with the parameter percentage_scribbled_regions = 0.1 in the config file

cd ./scribble_generation``
python create_dataframe_to_images.py --split train --percentage_scribbled_regions 0.1

This code generates healthy and tumorous scribbles for each slides of the CAMELYON 16 DATASET and save the scribbles as a list of coordinates (center of the patches along the scribbles) with the corresponding slide name and the corresponding class: healthy of tumor This code must be run on the two split: train and test

For each split, a dataset is then saved in PATH_CAMELYON

python dataframe_to_images.py --split train

For each split value , this code extract the images along the scribbles according to the dataframes created before and save them in the 2 folders: path_patches_scribbles_train and path_patches_scribbles_train. The class of the patch is written inside the filename of the corresponding patch.

2. Train VGG16 or ResNet on Scribbles

cd ./train_vgg
python train.py

Once the ResNet/VGG16 is trained, the weights are saved in path_weights and the filename correponding contains the value of percentage_scribbled_region.

3. Generate patches/predictions/ground truth values on the original test set

cd ./generate_patches_gts_preds
python generate_all_patches.py --k 0 --p 16

This code extract the tissue of each slide of the test set and create tiles with an overlap ov and save them afterwards. if you have num_processors = 16 and n_slides= 48 then the code must be ran 3 times with k = 0, 1, 2 The patches are saved in path_patches_test.

python monte_carlo.py --n_passes <number_mc_passes> --model <model> 

model can be either resnet50 or vgg16. This code uses the VGG16 trained in #2 and monte_carlo to extract for each patch of each slide:

1. n_passes predictions.
2. The feature vector.

the extracted data is save in path_prediction_features

python get_gt_patches.py

This code generate the ground truth values for each patch and save them to path_prediction_features.

4. generate heatmaps/uncertainty heatmaps

To generate a heatmap/uncerainty map:

cd ./heatmaps
python predict_and_create_heatmap_and_uncertainty_map.py --filename test_001 --uncertainty entropy

This 2 codes generate a heatmap for a given slide. To select the slide you can change the filename parameter. If parameter uncertainty is not set this code create a heatmap. For the uncertainty heatmap you can select one of the 2 uncertainty metric available in uncerainty metrics:(arg = 'entropy or 'std'):

1. compute_std
2. compute_entropy

The code extracts the patch, overlays the prediction on the patch, save the patch in patch_predictions_patches and stitch all the patches together and save the heatmap in path_uncertainty_maps or path_heatmaps.

5. SVM Corrections

cd ./correction_SVM
python generate_tables.py --split val --n_tables 10 --use_mc y

the code generate the tables in the paper ... you can either:

1. use monte-carlo or not for the correction process with the boolean use_mc = y/n.
2. Select the folder you want to do the correction process on : folder = val/test
3. Select the epochs range if you want to try different number of epochs for the correction step by changing epochs_range

If you use this code bellow please cite:

@misc{habis2024scribblebasedfastweaksupervisioninteractive,
      title={Scribble-based fast weak-supervision and interactive corrections for segmenting whole slide images}, 
      author={Antoine Habis and Roy Rosman Nathanson and Vannary Meas-Yedid and Elsa D. Angelini and Jean-Christophe Olivo-Marin},
      year={2024},
      eprint={2402.08333},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2402.08333}, 
}