Subsequent projects | Special projects

Start of funding 01.01.2021

Deep learning reconstruction with known operators for high-resolution CT

Prof. Dr. Andreas Maier
Friedrich-Alexander-University of Erlangen-Nuremberg
Computer Science Department 5 - Pattern Recognition Lab

Prof. Dr. Adam Wang
Stanford University
Radiological Sciences Laboratory


In general, the software pipelines that are able to learn generic end-to-end reconstruction models from input projections without explicitly modelling the imaging system have already been established by The group of Prof. Maier. This is achieved by applying supervised learning techniques to train a model formulated as an artificial neural network. However, as we have shown, the performance of the model is immensely improved by incorporating known operators. We propose to leverage the expert knowledge in hardware of imaging systems at Prof. Wang’s group in Stanford, to apply these known operators relating to spatial resolution to machine learning algorithms. This involves several steps. First, the imaging system will be evaluated in terms of focal spot size and shape, detector blur, and nominal acquisition geometry. The focal spot and detector blur are stable effects and will be incorporated as known operators. Second, the acquisition geometry will be refined since it is known to deviate from its nominal geometry due to scanner wobble and/or patient motion. The PYRO-NN framework will be used to learn the projection matrices that define the actual acquisition geometry. In addition, we will incorporate image regularization techniques that encourage natural-looking super resolution. The overall pipeline will first be developed, tested, and validated using simulated training data. We will then modify the existing algorithms, which are proven to be working on simulated data, to work on images acquired by Stanford’s CT imaging systems. The crucial point here will be the generation of training data, which will be tailored to the imaging system’s properties used in Stanford. We hypothesize that the combination will yield excellent high-resolution images from Stanford’s imaging systems and FAU’s novel machine learning algorithms.