Collaborative Computational Project in Synergistic Reconstruction for Biomedical Imaging

24th Software Meeting

Wednesday, July 24, 2019 - 13:30 to 16:00

Venue: University of Hull

Register to the event by 15/7/2019

Draft agenda:

  • Current status of SIRF and associated projects
  • Talks on using Machine Learning in image reconstruction (concentrating on software aspects)
  • Discussion on forthcoming Symposium on Synergistic Image Reconstruction (3-4 November) and the 2 associated training days (5-6 November)
  • Review of the SIRF long term plan
Speaker Title Abstract Affiliation
Casper da Costa-Luis Machine Learning for Image Reconstruction: Software Aspects   (Video) This talk discusses initial experience with integrating machine learning into an image reconstruction pipeline for PET. I'll include full installation instructions for Ubuntu, and examples using `NiftyPET` and `keras` (in Python 2). School of Biomed. Eng. & Im. Sci., KCL, St Thomas' Hospital, London SE1 7EH
Kuang Gong VideoSoftware aspects of applying deep learning to PET image reconstruction The combination of PET image reconstruction with deep learning, which includes the penalised reconstruction and unrolled reconstruction approaches, will be presented. In specific, implementation details, potential pitfalls and challenges across the process as a whole, will be introduced. In addition, optimisation directions and essential needs with regards to the platform and projectors will be discussed. Department of Radiology, Massachusetts General Hospital and Harvard Medical School.
Olivier Verdier Proof of concept: motion correction with deep learning We show how deep-learning based registration can be used to decrease noise by motion correction. The software architecture is based on the deep learning framework Voxelmorph (itself based on a layer of Keras atop of TensorFlow) for the registration, and the operator library ODL to compute the alternate MLEM algorithm. I will explain how these are all put together. KTH-Royal institute of technology (Stockholm, Sweden), HVL-Western Norway University of Applied Sciences (Bergen, Norway)
Christopher Syben PYRO-NN: Python Reconstruction Operators in Neural Networks The embedding of known operators in neural networks makes it possible to combine the power of deep learning with physics and signal processing. The software-related aspects of this approach are presented on the basis of the PYRO-NN framework. The PYRO-NN Framework brings CT reconstruction operators as CUDA-kernels to Tensorflow. We present implementation details and challenges we faced during the development of PYRO-NN. Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU)