Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Deep Learning Based Style Transfer for Improved Data Augmentation
Using style transfer for augmentation of the training data has been shown to help segmentation algorithms generalise better. We will apply these techniques to different imaging modalities and acquisition parameters of brain scans to improve disease detection algorithms.
Keywords: Brain imaging, Magnetic Resonance Imaging, Computed Tomography Imaging, Machine Learning, Neural Network, Style Transfer
The application of deep neural networks (DNNs) as disease detection algorithms has celebrated a number of important milestones towards clinical implementation in the recent past. However, these networks are usually trained on a limited data set, with data from a single (or few) institutions/scanners/modalities and certain acquisition parameters. In order for DNNs to be useful for clinical applications, the predictions must be robust across different domains of data. This is a problem known as domain adaptation.
In our group we apply DNNs to detect lesions in the brain (e.g. from ischemic stroke or multiple sclerosis) in a variety of imaging modalities. Using a large database of available brain scans, the aim of the project is to train a style transfer network, which will then be used as part of the data augmentation for the disease detection pipeline. Ultimately, this approach will be evaluated quantitatively against geometrical and other intensity-based data augmentation.
Further reading:
1. Application of intensity augmentation to breast segmentation: https://arxiv.org/pdf/1909.02642.pdf
2. Style transfer data augmentation: https://arxiv.org/pdf/1809.05375.pdf
Programming experience in Python and a basic understanding of neural networks is required. Prior experience with Tensorflow/Keras is also helpful, but not a prerequisite.
The application of deep neural networks (DNNs) as disease detection algorithms has celebrated a number of important milestones towards clinical implementation in the recent past. However, these networks are usually trained on a limited data set, with data from a single (or few) institutions/scanners/modalities and certain acquisition parameters. In order for DNNs to be useful for clinical applications, the predictions must be robust across different domains of data. This is a problem known as domain adaptation.
In our group we apply DNNs to detect lesions in the brain (e.g. from ischemic stroke or multiple sclerosis) in a variety of imaging modalities. Using a large database of available brain scans, the aim of the project is to train a style transfer network, which will then be used as part of the data augmentation for the disease detection pipeline. Ultimately, this approach will be evaluated quantitatively against geometrical and other intensity-based data augmentation.
Further reading: 1. Application of intensity augmentation to breast segmentation: https://arxiv.org/pdf/1909.02642.pdf 2. Style transfer data augmentation: https://arxiv.org/pdf/1809.05375.pdf
Programming experience in Python and a basic understanding of neural networks is required. Prior experience with Tensorflow/Keras is also helpful, but not a prerequisite.
The goals of this project are:
- Literature review on style transfer and basic MR theory.
- Prepare the database.
- Construct and train a style transfer network.
- Evaluate the results for cross-modality disease detection.
- Compare to other non-contrast specific intensity augmentation techniques.
- Write report and present results.
Details can be adapted according to the interests of the student.
The goals of this project are:
- Literature review on style transfer and basic MR theory. - Prepare the database. - Construct and train a style transfer network. - Evaluate the results for cross-modality disease detection. - Compare to other non-contrast specific intensity augmentation techniques. - Write report and present results.
Details can be adapted according to the interests of the student.
Dr. Moritz Platscher (platscher@biomed.ee.ethz.ch); Prof. Dr. Sebastian Kozerke. To apply for this project please attach a copy of your CV and transcripts of your Bachelor and Master studies.
Dr. Moritz Platscher (platscher@biomed.ee.ethz.ch); Prof. Dr. Sebastian Kozerke. To apply for this project please attach a copy of your CV and transcripts of your Bachelor and Master studies.