Data and Image Domain Deep Learning for Tomographic Computational Imaging

Wednesday, October 16, 2019 - 4:15pm - 5:00pm
Keller 3-180
W. Clem Karl (Boston University)
Computed Tomography is a non-invasive computational imaging modality with applications ranging from healthcare to security. It reconstructs cross-sectional images of an object using a collection of projection data collected at different angles. Conventional methods, such as FBP, require that the projection data be uniformly acquired over the complete angular range and of relatively high quality. Security is a domain where non-rotational scanning configurations are being developed which violate the complete data assumption and where conventional methods can produce images that are filled with artifacts. The recent success of deep learning methods in computer vision has inspired researchers to apply these powerful image-domain tools to tomographic problems with the aim of removing these image artifacts through post-processing. This approach has shown some success in certain applications, but can still leave perceptible residual artifacts. Another approach is to use deep learning in the data domain to pre-process and complete the original data, thus mitigating artifact creation during the image formation step. In this presentation, we present some of our work on data domain deep learning as well as recent extensions aiming to combine the power of both data domain and image domain learning through the consensus equilibrium framework. We show preliminary results obtained on challenging tomographic problems.