Preservational Self-supervised Learning

This repo is the official implementation of our ICCV 2021 paper titled "Preservational Learning Improves Self-supervised Medical Image Models by Reconstructing Diverse Contexts". In this repo, we demonstrate how to use PCLR to conduct pre-training on NIH ChestX-ray14 (2D) and LUNA (3D). The employed backbones are ResNet-18 and 3D U-Net, respectively. Note that this repo contains an improved version of our ICCV paper, which means it is possible to achieve higher results using codes in this repo. Also, we made some modifications, such as replacing the outer-product operation in transformation-conditioned attention with channel-wise multiplication, which results in more stable testing results.

Dependency

Please install PyTorch (>=1.1) before you run the code. We strongly recommend you to install Anaconda3 where we use Python 3.6. In addition, we use apex for acceleration. We also use pretrained-models.pytorch and segmentation_models_pytorch for convenience.

NIH ChestX-ray14 (Chest14)

Step 0

Please download Chest X-rays from this link.

The image folder of Chest14 should look like this:

./Chest14
	images/
		00002639_006.png
		00010571_003.png
		...

Besides, we also provide the list of training image in pytorch/train_val_txt/chest_train.txt.

Step 1

git clone https://github.com/Luchixiang/PCRL.git

cd PCRL/pytorch/

Step 2

python main.py --data chest14_data_path --phase pretask --model pcrl --b 64 --epochs 240 --lr 1e-3 --output pretrained_model_save_path --optimizer sgd --outchannel 3 --n chest --d 2 --gpus 0,1,2,3 --inchannel 3 --ratio 1.0

--data defines the path where you store Chest14.

--d defines the type of dataset, 2 stands for 2D while 3 denotes 3D.

--n gives the name of dataset.

--ratio determines the percentages of images in the training set for pretraining. Here, 1 means using all training images in the training set to for pretraining.

LUNA16

Step 0

Please download LUNA16 from this link

The image folder of LUNA16 should looks like this:

./LUNA16
	subset0		     		   	
		1.3.6.1.4.1.14519.5.2.1.6279.6001.979083010707182900091062408058.raw
		1.3.6.1.4.1.14519.5.2.1.6279.6001.979083010707182900091062408058.mhd
  	...
	subset1
	subset2
	...
	subset9

We also provide the list of training image in pytorch/train_val_txt/luna_train.txt.

Step1

git clone https://github.com/Luchixiang/PCRL.git

cd PCLR/pytorch

Step 2

First, you should pre-process the LUNA dataset to get cropped pairs from 3D images.

python preprocess/luna_pre.py --input_rows 64 --input_cols 64 --input_deps 32 --data LUNA_dataset_path --save processedLUNA_save_path

Step3

python main.py --data processed LUNA_save_path --phase pretask --model pcrl --b 16 --epochs 240 --lr 1e-3 --output pretrained_model_save_path --optimizer sgd --outchannel 1 --n luna --d 3 --gpus 0,1,2,3 --inchannel 1 --ratio 1.0