Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Resource-efficient 3D convolutional neural networks for high-resolution segmentation
The aim of this project is to develop a resource-efficient convolutional neural network for anatomical segmentation of three-dimensional magnetic resonance scans.
Keywords: convolutional neural networks, anatomical segmentation, deep learning
Anatomical segmentation of magnetic resonance imaging (MRI) scans is highly relevant in clinical diagnostics and scientific research. Segmentation maps are used for disease and anomaly detection and to derive quantitative information from a collection of scans.
In recent years learning-based segmentation systems, based on deep convolutional neural networks, have surpassed classical techniques in performance and processing speed. However, direct processing of three-dimensional MR volumes, with tens of millions of input voxels, is challenging due to memory constraints of graphical processing units.
During this project we will optimise an existing 3D convolutional neural network architecture in terms of its memory-efficiency and ideally increase the segmentation performance on MRI scans with up to 10 million input voxels. The optimisation may involve using neural architecture search techniques (https://arxiv.org/abs/1905.11946) or the integration of squeeze-and-excitation structures (https://https://arxiv.org/abs/1709.01507).
Interested candidates should have programming experience in Python and a basic understanding of convolutional neural networks. Prior experience with Tensorflow is also helpful, but can also be gained during the project.
Anatomical segmentation of magnetic resonance imaging (MRI) scans is highly relevant in clinical diagnostics and scientific research. Segmentation maps are used for disease and anomaly detection and to derive quantitative information from a collection of scans.
In recent years learning-based segmentation systems, based on deep convolutional neural networks, have surpassed classical techniques in performance and processing speed. However, direct processing of three-dimensional MR volumes, with tens of millions of input voxels, is challenging due to memory constraints of graphical processing units.
During this project we will optimise an existing 3D convolutional neural network architecture in terms of its memory-efficiency and ideally increase the segmentation performance on MRI scans with up to 10 million input voxels. The optimisation may involve using neural architecture search techniques (https://arxiv.org/abs/1905.11946) or the integration of squeeze-and-excitation structures (https://https://arxiv.org/abs/1709.01507).
Interested candidates should have programming experience in Python and a basic understanding of convolutional neural networks. Prior experience with Tensorflow is also helpful, but can also be gained during the project.
- Literature review on resource efficient convolutional neural network architectures.
- Implementation of network architectures and systematic investigation of the memory consumption.
- Training of the neural networks on existing databases of MRI scans and label maps.
- Evaluation of the segmentation performance.
- Literature review on resource efficient convolutional neural network architectures. - Implementation of network architectures and systematic investigation of the memory consumption. - Training of the neural networks on existing databases of MRI scans and label maps. - Evaluation of the segmentation performance.
Jonathan Zopes (zopes@biomed.ee.ethz.ch)
Moritz Platscher (platscher@biomed.ee.ethz.ch)
Sebastian Kozerke
Jonathan Zopes (zopes@biomed.ee.ethz.ch) Moritz Platscher (platscher@biomed.ee.ethz.ch) Sebastian Kozerke