Post Doctoral Fellowship on Deep Learning and Clinical Spectral-CT
The research group in Mathematical Imaging within the Department of Mathematics is offering a two-year postdoctoral fellowship based on a grant from the applied mathematics programme at the Swedish Foundation for Strategic Research.
The position is part of a larger medical imaging project where the overall goal is to develop theory and algorithms for image reconstruction applicable to x-ray based medical imaging with under-sampled and/or highly noisy data. Overall clinical goals are to significantly reduce the total dose of x-rays and/or acquisition time while maintaining a clinically useful image quality, alternatively to significantly improve image quality given a fixed total dose/acquisition time. Imaging modalities involved are 3D spiral/helical computed tomography (CT), 3D spatiotemporal SPECT/CT and PET/CT, C-arm 3D-CT, and spectral-CT. The project also involves applications to x-ray and electron microscopy phase contrast imaging.
The position includes research into theory and development of algorithms that use methods from machine learning for image reconstruction in spectral-CT. The work also includes implementation of algorithms that can be applied to clinical data from stroke and/or oncological imaging. A key element is to design deep neural network architectures for image reconstruction that incorporate an explicit physics model for how data is generated in spectral-CT as well as a statistical characterization of the noise. Another key element is to handle multichannel data and images, especially bearing in mind the limited memory in GPU-hardware. The large-scale nature of the reconstruction problem in spectral-CT coupled with the time-critical nature of some of its clinical applications, requires algorithms that not only execute fast but also minimizes memory footprint. Prototype algorithms will be implemented as software components in ODL (http://github.com/odlgroup/odl) using a suitable deep learning framework, like TensorFlow, as back-end. ODL is a Python-based software framework for numerical functional analysis that the research group in Mathematical Imaging uses for prototyping.
Much of the research will be pursued at MedTechLabs (https://www.medtechlabs.se) and at the Medical Imaging group at the Department of Physics, KTH. The group, which is led by Prof. Mats Danielsson, has developed novel photon-counting silicon strip detectors for clinical CT-scanners that are now entering tests in a clinical setting at MedTechLabs.
This is a unique opportunity to pursue mathematical and algorithmic research to address challenges that arise when photon counting spectral-CT imaging is used in a clinical setting. As a postdoctoral fellow, you will have access to unique clinical data and from one of the first clinical spectral-CT scanners. You will also benefit from the strong research environments at KTH in mathematical sciences and medical imaging physics. Finally, through MedTechLabs you also gain access the essential clinical expertise.
We seek a candidate with a PhD degree in mathematics, signal processing or computational physics/engineering that has been awarded (or planned to be awarded) before the commencement of the position. The candidate should have a strong background from machine learning or signal/image processing, the latter preferably in the context of tomographic image reconstruction. Experience from computational harmonic analysis, functional analysis, and statistical learning is highly beneficial. The candidate should also have experience from software development in scientific computing, preferably using Python and/or C/C++ in the context of machine learning. Finally, a successful applicant must be strongly motivated and have the capability to work independently as well as in collaboration with members of the research group.
Deadline: March 1, 2019