Unsupervised domain adaptation for clinician pose estimation and instance segmentation in the operating room

Vinkle Srivastav, Afshin Gangi, Nicolas Padoy, Medical Image Analysis (2022)

AdaptOR

Installation requirements

Python ≥ 3.6, PyTorch ≥ 1.5, and torchvision with matching PyTorch version.
Detectron2: follow the INSTALL.md to install detectron2.
Gradio: for the demo pip install gradio
We used python=3.7.8, pytorch=1.6.0, torchvision=0.7, detectron2=0.3, cuda=10.2, and cudnn=7.6.5 in our experiments.
Run the following commands to download the repo.

($)> AdaptOR=/path/to/the/repository
($)> git clone https://github.com/CAMMA-public/adaptor.git $AdaptOR
($)> cd $AdaptOR

Datasets

Follow the Datasets.md to download the COCO dataset. The directory structure should look as follows.

adaptor/
└── datasets/
    └── coco/
        ├── train2017/
        ├── val2017/
        └── annotations/
        	├── person_keypoints_train2017.json
        	└── person_keypoints_val2017.json
    ...

We will soon release the TUM-OR dataset needed for the domain adaptation experiments for AdaptOR

Model Weights for the AdaptOR-SSL experiments

Model	Supervision	AP^bb_person	AP^kp_person	AP^mask_person	Model Weights
MoCo-v2_R_50_detectron2	-	-	-	-	download
kmrcnn_R_50_FPN_3x_gn_amp_sup1	1%	42.57	38.22	36.06	download
kmrcnn_R_50_FPN_3x_gn_amp_sup2	2%	45.37	44.08	38.96	download
kmrcnn_R_50_FPN_3x_gn_amp_sup5	5%	49.90	49.79	43.10	download
kmrcnn_R_50_FPN_3x_gn_amp_sup10	10%	52.70	56.65	45.46	download
kmrcnn_R_50_FPN_3x_gn_amp_sup100	100%	57.50	66.20	49.90	download

Evaluation

First, download the pre-trained models and copy inside the $AdaptOR/models. Run the eval_ssl_coco.sh or see the following bash script for the evaluation of AdaptOR-SSL on the COCO dataset with different percentages of supervision.

PERCENTAGE_SUPERVISION=1 #2,5,10,100
CONFIG_FILE=./configs/ssl_coco/kmrcnn_R_50_FPN_3x_gn_amp_sup${PERCENTAGE_SUPERVISION}.yaml
MODEL_WEIGHTS=./models/kmrcnn_R_50_FPN_3x_gn_amp_sup${PERCENTAGE_SUPERVISION}.pth
python train_net.py \
      --eval-only \
      --config ${CONFIG_FILE} \
      MODEL.WEIGHTS ${MODEL_WEIGHTS}

Training

Run the train_ssl_coco.sh or see the following bash script for the training of AdaptOR-SSL on the COCO dataset with different percentages of supervision on 4 V100 GPUS. To train on more or less GPUs, follow the linear scaling rule.

PERCENTAGE_SUPERVISION=1 #2,5,10,100
CONFIG_FILE=./configs/ssl_coco/kmrcnn_R_50_FPN_3x_gn_amp_sup${PERCENTAGE_SUPERVISION}.yaml
python train_net.py \
      --num-gpus 4 \
      --config ${CONFIG_FILE} \

Training on the TUM-OR dataset for the domain adaptation experiments for AdaptOR coming soon.

Gradio Demo

Download the pre-trained models and copy inside the $AdaptOR/models

($)> python AdaptOR-SSL_demo_image.py

Demo outputs

Sample output of AdaptOR comparing the model before and after the domain adaption on the downsampled images with downsampling factor of 8x (80x60)

Sample output of AdaptOR-SSL comparing the model trained with 1% vs. 100% supervision

Citation

@inproceedings{srivastav2021adaptor,
  title={Unsupervised domain adaptation for clinician pose estimation and instance segmentation in the operating room},
  author={Srivastav, Vinkle and Gangi, Afshin and Padoy, Nicolas},
  booktitle={Medical Image Analysis},
  year={2022},
}

References

The project uses detectron2. We thank the authors of detectron2 for releasing the library. If you use detectron2, consider citing it using the following BibTeX entry.

@misc{wu2019detectron2,
  author =       {Yuxin Wu and Alexander Kirillov and Francisco Massa and
                  Wan-Yen Lo and Ross Girshick},
  title =        {Detectron2},
  howpublished = {\url{https://github.com/facebookresearch/detectron2}},
  year =         {2019}
}

The project also leverages following research works. We thank the authors for releasing their codes.

Unbiased-Teacher
Rand-Augment and Rand-Cut
Pytorch Image Models
MoCo-v2 pre-trained weights are used to initialize the AdaptOR-SSL models. The MoCo-v2 weights in the detectron2 format is available here MoCo-v2-detectron2. The config files by default download the weights from our S3 server. You can also use this script to convert the MoCo-v2 weights to the detectron2 format.

License

This code, models, and datasets are available for non-commercial scientific research purposes as defined in the CC BY-NC-SA 4.0. By downloading and using this code you agree to the terms in the LICENSE. Third-party codes are subject to their respective licenses.

CAMMA-public / HPE-AdaptOR