CA2CL: Cluster-Aware Adversarial Contrastive Learning for Pathological Image Analysis

This is a PyTorch implementation of the CA2CL

Preparation

Download the NCT-CRC-HE-100K and CRC-VAL-HE-7K：NCT dataset.

Unsupervised Training

To do unsupervised pre-training of a ResNet-18 model on NCT-CRC-HE-100K in an 2*A100(40G) gpu machine, run:

CUDA_VISIBLE_DEVICES=0,1 python pretrain.py 

the NCT pretrained checkpoint is available at checkpoint

Fine-tuning

With a pre-trained model, to train a supervised linear classifier on frozen features/weights in an 1-gpu machine, run:

python ft_le.py --dataset_path '../data/Kather_Multi_Class/' \
                --model_name 'resnet18' \
                --model_path './save/NCT/pretrain/resnet18/CA2CL_NCT.tar' \
                --ft_epoch 40 \ 
                --lr 0.0003 \ 
                --seed 1 \ 
                --only_fine_turning \ 
                --labeled_train 0.01 \ 
                --gpu_index 3

Fine-tuning results on CRC-VAL-HE-7K (seed=1):

	label of train 1%	label of train 10%	label of train 50%	label of train 100%
acc	0.958	0.967	0.968	0.970
f1	0.941	0.954	0.957	0.960
logs	download	download	download	download

Linear Classification

With a pre-trained model, to train a supervised linear classifier on frozen features/weights in an 8-gpu machine, run:

python ft_le.py --dataset_path '../data/Kather_Multi_Class/' 
                --model_name 'resnet18'
                --model_path './save/NCT/pretrain/resnet18/CA2CL_NCT.tar' \
                --ft_epoch 100
                --lr 0.01
                --seed 1
                --only_linear_eval
                --labeled_train 0.01
                --gpu_index 3

Linear classification results on CRC-VAL-HE-7K (seed=1):

	label of train 1%	label of train 10%	label of train 50%	label of train 100%
acc	0.967	0.969	0.969	0.968
f1	0.953	0.959	0.958	0.956
logs	download	download	download	download

Transferring to Object Detection

We follow the evaluation setting in MoCo when trasferring to object detection.

1. Install detectron2..

2. You can download our pre-processed datasets from GlaS-coco-format and CRAG-coco-format.

3. Put dataset under "benchmarks/detection/datasets" directory, following the directory structure requried by detectron2.

4. the converted backbone weights is available at checkpoint, put dataset under "benchmarks\detection\converted_weights" directory,

5. run run_ft.sh runs of fine-tuning and evaluation on GlaS dataset.

License

This project is under the CC-BY-NC 4.0 license. See LICENSE for details.