Standard view images are important in clinical practice as they provide a means to perform biometric measurements from similar anatomical regions. In this project, we employ a multi-scale reinforcement learning (RL) agent framework that enables a natural learning paradigm by interacting with the environment and mimicking experienced operators' navigation steps.
Here are few examples of the learned agent for plane detection on unseen data:
- Detecting the 4-chamber plane in short-axis cardiac MRI acquisition (HQ video)
- Detecting the axial plane containing the anterior and posterior commissure (ACPC) line in adult brain MRI acquisition (HQ video)
python DQN.py --algo DQN --gpu 0
python DQN.py --algo DQN --gpu 0 --task play --load path_to_trained_model
If you use this code in your research, please cite this paper:
@inproceedings{alansary2018automatic,
title={Automatic view planning with multi-scale deep reinforcement learning agents},
author={Alansary, Amir and Le Folgoc, Loic and Vaillant, Ghislain and Oktay, Ozan and Li, Yuanwei and Bai, Wenjia and Passerat-Palmbach, Jonathan and Guerrero, Ricardo and Kamnitsas, Konstantinos and Hou, Benjamin and others},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={277--285},
year={2018},
organization={Springer}
}