Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Towards cross-scanner robustness in deep learning based medical image segmentation
Neural Networks lack robustness to out-of-distribution samples. We explore different domain generalization solutions to tackle this problem in the context of medical image segmentation.
Keywords: medical image segmentation, cross-scanner robustness, deep learning, domain generalization.
Neural Networks demonstrate state-of-the-art performance on several medical image analysis tasks, but often lack robustness to out-of-distribution samples. This situation presents itself quite frequently in medical imaging - for instance, a neural network trained on images from one MRI scanner suffers drastic performance degradation when tested on images acquired from another scanner. This is one of the main hurdles obstructing large scale clinical utilization of these methods. The desired behavior that would be a solution to this problem is referred to as 'domain generalization' in the machine learning literature - meaning that an algorithm trained on one or more training domains (scanners / hospitals / imaging protocols) then generalizes to unseen domains that are unknown during training. Several approaches have been suggested for domain generalization (see reference list), most of which have tackled the problem in the context of image classification. In this project, we aim to obtain an extensive understanding of existing approaches and to extend suitable approaches to medical image segmentation. For experimentation, we will use multi-scanner datasets for 3 different anatomies - namely, the brain, heart and prostate.
References:
[1] Volpi, Riccardo, et al. "Generalizing to unseen domains via adversarial data augmentation." Advances in Neural Information Processing Systems. 2018.
[2] Li, Da, et al. "Learning to generalize: Meta-learning for domain generalization." Thirty-Second AAAI Conference on Artificial Intelligence. 2018.
[3] Li, Haoliang, et al. "Domain generalization with adversarial feature learning." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.
[4] Dou, Qi, et al. "Domain generalization via model-agnostic learning of semantic features." Advances in Neural Information Processing Systems. 2019.
[5] Wang, Haohan, et al. "Learning robust global representations by penalizing local predictive power." Advances in Neural Information Processing Systems. 2019.
Neural Networks demonstrate state-of-the-art performance on several medical image analysis tasks, but often lack robustness to out-of-distribution samples. This situation presents itself quite frequently in medical imaging - for instance, a neural network trained on images from one MRI scanner suffers drastic performance degradation when tested on images acquired from another scanner. This is one of the main hurdles obstructing large scale clinical utilization of these methods. The desired behavior that would be a solution to this problem is referred to as 'domain generalization' in the machine learning literature - meaning that an algorithm trained on one or more training domains (scanners / hospitals / imaging protocols) then generalizes to unseen domains that are unknown during training. Several approaches have been suggested for domain generalization (see reference list), most of which have tackled the problem in the context of image classification. In this project, we aim to obtain an extensive understanding of existing approaches and to extend suitable approaches to medical image segmentation. For experimentation, we will use multi-scanner datasets for 3 different anatomies - namely, the brain, heart and prostate.
References: [1] Volpi, Riccardo, et al. "Generalizing to unseen domains via adversarial data augmentation." Advances in Neural Information Processing Systems. 2018. [2] Li, Da, et al. "Learning to generalize: Meta-learning for domain generalization." Thirty-Second AAAI Conference on Artificial Intelligence. 2018. [3] Li, Haoliang, et al. "Domain generalization with adversarial feature learning." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018. [4] Dou, Qi, et al. "Domain generalization via model-agnostic learning of semantic features." Advances in Neural Information Processing Systems. 2019. [5] Wang, Haohan, et al. "Learning robust global representations by penalizing local predictive power." Advances in Neural Information Processing Systems. 2019.
In this project, the student will:
- familiarize themselves with related literature.
- implement suitable algorithms in a deep learning framework such as Tensorflow, PyTorch, etc.
- assess / improve the accuracy of existing methods and extend the methods for further analysis.
- if successful, publish their findings.
In this project, the student will: - familiarize themselves with related literature. - implement suitable algorithms in a deep learning framework such as Tensorflow, PyTorch, etc. - assess / improve the accuracy of existing methods and extend the methods for further analysis. - if successful, publish their findings.
- Neerav Karani (nkarani@vision.ee.ethz.ch)
- Dr. Ertunc Erdil (ertunc.erdil@vision.ee.ethz.ch)
- Prof. Ender Konukoglu (kender@vision.ee.ethz.ch)
- Neerav Karani (nkarani@vision.ee.ethz.ch) - Dr. Ertunc Erdil (ertunc.erdil@vision.ee.ethz.ch) - Prof. Ender Konukoglu (kender@vision.ee.ethz.ch)