Interventional Multi-Instance Learning with Deconfounded Instance-Level Prediction

Implementation of IMIL.

Environments

• OS: Ubuntu18.04.4 LTS
• GCC/G++: 7.5.0
• GPU: GeForce GTX TITAN X *1 |12G Memory
• CUDA Version: 10.1
• NVIDIA-SMI: 450.57
• CPU: Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz * 12.

Installation

Install requirements:

pip install -r requirements.txt

For faiss, we recommend installing with the following command:

pip install faiss-gpu

See faiss-wheels for more details.

For PyTorch , we recommend following the instruction on official website.

Data Folder Structure

We support training and testing on DigestPath, Camelyon16and Pascal VOC 07. Their data roots are like:

Data Root/
	DigestPath/
		five_folder/
			f0/
				train/
					pos/
						WSI_0/
							WSI_O_X_Y_(instance_label,bag_label).png
				        neg/
				test/
            	
	Camenlyon16/
		train/
			pos_instances_group_in_pos_bags/
				WSI_0/
				WSI_3/
   		 	neg_instances_group_in_pos_bags/
				WSI_0/
				WSI_3/
            		neg_instances_group_in_neg_bags/
				WSI_1/
                
		test/
		
	Pascal VOC 2007/
		JPEGImages/
		    Bag_0/
		    Bag_1/
		Main/
			ImageSets/
				bird_train.txt
				bird_test.txt
				bird_trainval.txt
				bird_val.txt
				...

Reproduce DigestPath

This implementation supports single-gpu training; multi-gpu, DistributedDataParallel is TDB.

To do multiple instance learning with a ResNet-18 model in single-gpu machine:

python main.py  --task DigestSeg --config DigestSegEMCAV2 \
	        --log_dir  [your logging folder] \
	        --data_root  [your data root folder] \
	        --workers 8  --ignore_thres 0.95 --mmt 0.5 

Reproduce Camelyon16

Please run:

python main.py --task Camelyon --config DigestSegEMCAV2 \
	       --log_dir [your logging folder] \
	       --data_root [your data root folder] \
	       --workers 8 -lr 1e-3 --epochs 50 --mmt 0.5 --ignore_thres 0.95 

Reproduce Pascal VOC 07

Please run:

python main.py --task Pascal --config DigestSegEMCAV2 \
	       --log_dir [your logging folder] \ 
               --workers 8 --pretrained --backbone 50 --ignore_thres 0.95 --mmt 0.5

Methods

The reproduced methods by ourselves including Oracle, RCEMIL, PatchCNN, SemiMIL, SimpleMIL and Top-kMIL. To run these methods, you should choose the corresponding configs.

To run the Oracle model on DigestPath:

python main.py  --task DigestSeg --config DigestSegFull \
		--log_dir  [your logging folder] \  
	        --data_root  [your data root folder] \
	        --workers 8 

To run the RCEMIL model on DigestPath:

python main.py  --task DigestSeg --config DigestSegRCE \
		--log_dir  [your logging folder] \  
	        --data_root  [your data root folder] \
	        --workers 8 --mmt 0.75 

To run the PatchCNN model on DigestPath:

python main.py  --task DigestSeg --config DigestSegPB \
		--log_dir  [your logging folder] \  
	        --data_root  [your data root folder] \
		--workers 8  --mmt 0 

To run the SemiMIL model on DigestPath:

python main.py  --task DigestSeg --config  DigestSemi \
		--log_dir  [your logging folder] \  
	        --data_root  [your data root folder] \
		--workers 8 --semi_ratio 0.5

To run the SimpleMIL model on DigestPath:

python main.py  --task DigestSeg --config  DigestSeg \
		--log_dir  [your logging folder] \  
	        --data_root  [your data root folder] \
		--workers 8 

To run the Top-kMIL model on DigestPath:

python main.py  --task DigestSeg --config DigestSegTOPK \
		--log_dir  [your logging folder] \  
	        --data_root  [your data root folder] \
		--workers 8 --mmt 0 

License

This project is under the MIT license. See LICENSE for details.