Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Many-Organ Segmentation
There is considerable existing work on using machine learning to automatically segment the heart in magnetic resonance (MR) images. However, there is little work on simultaneously segmenting the surrounding anatomy. This project is about exploring approaches to overcome this problem.
With sufficient high-quality training data, machine learning methods can achieve very impressive segmentation results, in some cases equaling human accuracy. However, without the availability of (sufficient) training data, the segmentation task becomes much less straightforward. For this reason, it is difficult to train neural networks to segment multiple organs from magnet resonance (MR) images, as creating training data would requirer manually labelling all of the desired organs in a large number of images, which would take considerable time.
However, if we could automatically segment multiple organs in MR images, then this would allow large numbers of images to be quicklly segmented, producing a range of examples of real anatomy. The resulting segmentation data would allow modelling the seen morphological diversity of anatomy, and the learned model could then be used to generate realistic in-silico phantoms.
Thus, it is desirable to develop methods to segment multiple organs in MR images, without relying on large quantities of labeled data. In this project the aim would be to explore whether unpaired image-to-image translation models (such as CycleGAN) could allow multi organ segmentation to be achieved in an unsupervised way.
With sufficient high-quality training data, machine learning methods can achieve very impressive segmentation results, in some cases equaling human accuracy. However, without the availability of (sufficient) training data, the segmentation task becomes much less straightforward. For this reason, it is difficult to train neural networks to segment multiple organs from magnet resonance (MR) images, as creating training data would requirer manually labelling all of the desired organs in a large number of images, which would take considerable time.
However, if we could automatically segment multiple organs in MR images, then this would allow large numbers of images to be quicklly segmented, producing a range of examples of real anatomy. The resulting segmentation data would allow modelling the seen morphological diversity of anatomy, and the learned model could then be used to generate realistic in-silico phantoms.
Thus, it is desirable to develop methods to segment multiple organs in MR images, without relying on large quantities of labeled data. In this project the aim would be to explore whether unpaired image-to-image translation models (such as CycleGAN) could allow multi organ segmentation to be achieved in an unsupervised way.
The aim of this project is to develop a segmentation network capable of accurately segmenting multiple organs and regions (e.g. heart, lungs, liver, ribs, fat, spine, etc) for MR images. If that stage is successfully achieved, there is then potential to explore generative models of the anatomical diversity seen in real data.
The aim of this project is to develop a segmentation network capable of accurately segmenting multiple organs and regions (e.g. heart, lungs, liver, ribs, fat, spine, etc) for MR images. If that stage is successfully achieved, there is then potential to explore generative models of the anatomical diversity seen in real data.
primary supervisor: Thomas Joyce joycet@ethz.ch,
Prof. Dr. Sebastian Kozerke
primary supervisor: Thomas Joyce joycet@ethz.ch, Prof. Dr. Sebastian Kozerke