Efficient Cancer Classification by Coupling Semi Supervised and Multiple Instance Learning

This is the code of the paper A. Schmidt, J. Silva-Rodríguez, R. Molina and V. Naranjo, "Efficient Cancer Classification by Coupling Semi Supervised and Multiple Instance Learning," in IEEE Access, vol. 10, pp. 9763-9773, 2022, doi: 10.1109/ACCESS.2022.3143345.

We try our best to make the code reusable and the experiments reproducible by giving a detailed instruction, description of dependencies, configurations and run commands:

Installation and Usage

To make this code run on your linux machine you need to:

• Install miniconda (or anaconda): https://docs.anaconda.com/anaconda/install/linux/
• Set up a conda environment and activate it:
  • conda env create --file environment.yaml
  • conda activate tensorlfow_2_3
• Download dataset, see dataset_dependent folders READMEs:
  • ./dataset_dependent/camelyon16/README.md
  • ./dataset_dependent/sicapv2/README.md
• Edit the configuration:
  • ./config.yaml for general settings
  • ./dataset_dependent/sicapv2/config.yaml for dataset dependent settings
• Run the program:
  • python ./src/main.py

Reproduce the Experiments of the Paper

In the paper we show the following experiment results for patch-level Gleason grading (on SICAPv2) and WSI-level breast cancer classification (on Camelyon16):

To run the experiments, please follow this instructions:

Patch-level Gleason Grading of Prostate Cancer (SICAPv2)

1. Follow the steps above to install dependencies and download the dataset
2. Configure the path to the SICAPv2 dataset on your Pc:
   • open ./dataset_dependent/sicapv2/config.yaml
   • change the line dir: path/to/dataset/
3. Run the experiments
   • Navigate into the base folder (cancer_classification)
   • The subfolders efficient_labeling and complete_annotation of ./dataset_dependent/sicapv2/experiments/ contain the configurations of the experiments in Figure 2 of the paper.
   • To train the model with efficient labeling (EL) and P=5 use f.e.:
     • python src/main.py -dc ./dataset_dependent/sicapv2/experiments/efficient_labeling/P_5/config.yaml
   • To test the model use the test configurations test_config.yaml, f.e.:
     • python src/main.py -dc ./dataset_dependent/sicapv2/experiments/efficient_labeling/P_5/test_config.yaml
4. To see the output, see below description of logging

WSI-level classification of Breast Cancer (Camelyon16)

1. Follow the steps above to install dependencies and download the dataset
   • Preprocessing is necessary, see dataset_dependent/camelyon16/dataset_scripts/README.md
2. Configure the path to the preprocessed Camelyon16 dataset on your Pc:
   • open ./dataset_dependent/camelyon16/config.yaml
   • change both paths dir: /path/to/cam16 and data_split_dir: /path/to/cam16 to the path of thepreprocessed dataset
3. Run the experiments with efficient labeling:
   • Navigate into the base folder (cancer_classification)
   • To train the model use the configurations of the subfolders of ./dataset_dependent/camelyon16/experiments/ f.e.:
     • python src/main.py -dc ./dataset_dependent/camelyon16/experiments/efficient_labeling/P_5/config.yaml
   • To test the model use the test configurations test_config.yaml, f.e.:
     • python src/main.py -dc ./dataset_dependent/camelyon16/experiments/efficient_labeling/P_5/test_config.yaml
4. To see the output, see below description of logging.

Experiment Logging

• Logging is done with mlflow (https://mlflow.org/docs/latest/tracking.html), it is already installed if you followed the installation guidlines above
• To see the experiment results, navigate into the base folder (cancer_classification). If you ran experiments, a mlruns folder should be present.
• Run mlflow ui
• Open in your browser localhost:5000 to see the results (training progress, metrics etc.)
• The models of the experiments by default are stored in the experiment subfolders, f.e. ./dataset_dependent/sicapv2/experiments/efficient_labeling/P_5/models/