Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Master’s Thesis: A Spine Segmentation Pipeline for MR Images using Unsupervised Domain Adaptation
There exists an accurate and robust spine segmentation pipeline for CT images (anduin.bonescreen.de). This is a series of three networks as shown in the figure (left), that are trained in a supervised manner using thousands of annotated CT scans. Can we now transfer these networks to segment MR images without any ground truth annotations?
Keywords: Deep Learning, Medical Image Analysis, Machine Learning, Computer Vision, Biomedical Image Analysis
There exists an accurate and robust spine segmentation pipeline for CT images (anduin.bonescreen.de). This is a series of three networks as shown in the figure (left), that are trained in a supervised manner using thousands of annotated CT scans1. Can we now transfer these networks to segment MR images without any ground truth annotations? The aim of this project is to use and develop novel machine learning algorithms to perform unsupervised domain adaptation, eventually repurposing every network in the ‘anduin’ pipeline to process MR images. This involves: (1) Building on state-of-art 2D domain adaptation methods to work efficiently on 3D medical images (2) Deploy the final pipeline at scale, > 10000 MR scans (3) Work with medical experts to finalise MR annotations.
There exists an accurate and robust spine segmentation pipeline for CT images (anduin.bonescreen.de). This is a series of three networks as shown in the figure (left), that are trained in a supervised manner using thousands of annotated CT scans1. Can we now transfer these networks to segment MR images without any ground truth annotations? The aim of this project is to use and develop novel machine learning algorithms to perform unsupervised domain adaptation, eventually repurposing every network in the ‘anduin’ pipeline to process MR images. This involves: (1) Building on state-of-art 2D domain adaptation methods to work efficiently on 3D medical images (2) Deploy the final pipeline at scale, > 10000 MR scans (3) Work with medical experts to finalise MR annotations.
Create an MR spine segmentation pipeline and open-source the segmentation tool to the research community.
Create an MR spine segmentation pipeline and open-source the segmentation tool to the research community.