Talk on Deep learning and Tomography.
Machine learning and machine vision methods have been used in conjunction with tomography for a long time, in relation for example to image segmentation, particularly in the medical field There has been signi cant recent progress in the field of machine learning with the advent of deep learning methods. Together with the associated scientic progress embodied in these methods, new frameworks have been proposed to actually implement and make these new methods more easily available, particularly with the TensorFlow and Torch/Pytorch packages. While it is not entirely obvious how machine learning techniques can be integrated with tomographic re- construction, these software packages can essentially be seen as combining linear and non-linear operators with effective and effcient optimisation methods implemented on GPUs. Seen in this light, we show how recent Deep Learning software packages can be repurposed to be used for efficiently computing projection and back-projection operators, regularization kernels and variational frameworks for joint tomographic reconstruction and denoising and segmentation. Of particular interest are the automated differentiation facilities in these packages , which simplify access to state-of-the-art inverse problem solving methods, as well as potientially accelerate them significantly.
The slides are here