DRAS-MIL

Discriminative Region Active Sampling for Multiple Instance Learning

Preprint | Presentation & Conference Paper | Cite

DRAS-MIL is a sampling approach to improve the efficiency of evaluating histopathology slides while minimising the loss of classification accuracy.

Workflows

Baseline model training

1. Whole slide tissue detection and patching (create_patches_fp.py)
2. Feature extraction (extract_features_fp.py)
3. Creation of cross-validation folds (create_splits_seq.py)
4. Model training (main.py)
5. Slide evaluation (eval.py)
6. Model evaluation (other_metrics.py)

DRAS-MIL evaluation experiments

Here we pre-compute all features as we will run multiple experiments, so will not save time by only computing relevant features

1. Whole slide tissue detection and patching (create_patches_fp.py)
2. Feature extraction (extract_features_fp.py)
3. Slide evaluation (eval.py with --sampling)
4. Model evaluation (other_metrics.py)

DRAS-MIL evaluation in practice

Here features are evaluated only when needed

1. Whole slide tissue detection and patching (create_patches_fp.py)
2. Slide evaluation (eval.py with --sampling and --eval_features)
3. Model evaluation (other_metrics.py)

Example code runs

Create 3 cross-validation folds:

python create_splits_seq.py --task custom_714 --seed 1 --label_frac 1 --val_frac 0.33 --test_frac 0.33 --k 3

Baseline model training with 500 random hyperparameter tuning experiments:

python main.py --coords_path "../../../../MULTIX/DATA/coords" --tuning --no_inst_cluster --num_tuning_experiments 500 --tuning_output_file tuning_results/main_custom1vsall_714_ABMILsb_ce_finaltuning.csv --split_dir /workspace/CLAM-private/splits/custom_714_100 --k 1 --results_dir /workspace/CLAM-private/results --exp_code main_custom1vsall_714_ABMILsb_ce_finaltuning --weighted_sample --bag_loss ce --task ovarian_1vsall --min_epochs 50 --max_epochs 500 --model_type clam_sb --log_data --subtyping --data_root_dir "/MULTIX/DATA/" --csv_path 'dataset_csv/set_all_714.csv' --features_folder "ovarian_dataset_features_256_patches_20x"

Baseline model training with the best hyperparameters found during tuning:

python main.py --early_stopping --use_all_samples --no_inst_cluster --reg 0.00079 --drop_out 0.02 --lr 0.0038 --split_dir /workspace/CLAM-private/splits/custom_714_100 --k 3 --results_dir /workspace/CLAM-private/results --exp_code main_nosampling_reg00079_dropout02_lr0038_1vsall_714_ABMILsb_ce_last10best_mean20stopper --weighted_sample --bag_loss ce --task ovarian_1vsall --max_epochs 500 --model_type clam_sb --log_data --subtyping --data_root_dir "/MULTIX/DATA/" --csv_path 'dataset_csv/set_all_714.csv' --features_folder "ovarian_dataset_features_256_patches_20x"

Slide evaluation processing all possible patches:

python eval.py  --drop_out 0.02 --split val --splits_dir /workspace/CLAM-private/splits/custom_714_100 --k 3 --models_exp_code main_nosampling_reg00079_dropout02_lr0038_1vsall_714_ABMILsb_ce_last10best_mean20stopper_s1 --save_exp_code main_nosampling_reg00079_dropout02_lr0038_1vsall_714_ABMILsb_ce_last10best_mean20stopper_VALIDSET --task ovarian_1vsall --model_type clam_sb --results_dir /workspace/CLAM-private/results/ --data_root_dir "/MULTIX/DATA/" --csv_path 'dataset_csv/set_all_714.csv' --features_folder "ovarian_dataset_features_256_patches_20x"

Bootstrapping to estimate model performance:

python bootstrapping.py --num_classes 2 --model_names main_nosampling_reg00079_dropout02_lr0038_1vsall_714_ABMILsb_ce_last10best_mean20stopper_VALIDSET --bootstraps 1000 --run_repeats 1 --folds 3

Reference

This code is forked from the CLAM repository with corresponding paper. This repository and the original CLAM repository are both available for non-commercial academic purposes under the GPLv3 License.

Breen, J., Allen, K., Zucker, K., Hall, G., Orsi, N.M. and Ravikumar, N., 2023, April. Efficient subtyping of ovarian cancer histopathology whole slide images using active sampling in multiple instance learning. In Proceedings of SPIE 12471 (Vol. 12471). SPIE. https://doi.org/10.1117/12.2653869.

@inproceedings{breen2023efficient,
  title={Efficient subtyping of ovarian cancer histopathology whole slide images using active sampling in multiple instance learning},
  author={Breen, Jack and Allen, Katie and Zucker, Kieran and Hall, Geoff and Orsi, Nicolas M and Ravikumar, Nishant},
  booktitle={Proceedings of SPIE 12471},
  volume={12471},
  year={2023},
  organization={SPIE}
}