Motion Recovery from Radon Transformed Image Using Neural Networks

Haoran Chang, Debasis Mitra

Keywords: motion correction, CNN, convolutional enoder decoder, inverse radon transform

TL;DR: We recover convolution function for motion blur and reconstruct image directly from sinograms using neural netowrks.

Abstract: In this paper we address motion correction in image reconstruction. Patient motion, including breathing, is a persistent problem in medical imaging. Cardiac motion is particularly enigmatic and often ignored, except in high-speed imaging modalities like MRI. Motions may create artifacts to the extent that the image may have to be discarded. Since the beginning of medical imaging, motion correction remained an important subcategory of research. Motion corrections may be applied during or after tomographic image reconstruction. In this work we considered motion as a Gaussian blur at the image level. Discrete radon transform is applied to the blurred images to create corresponding noisy sinograms that mimic real imaging scenario. Our deep learning based tool recovered accurately (1) the blurring functions with an artificial convolutional neural network (CNN) directly from the sinograms, and (2) successfully reconstructed (inverse radon transformed) the noise-free images utilizing an adaptation of the convolutional encoder-decoder network (CED) from the literature. Our work shows that neural networks are not only capable of eliminating systematic noise in reconstruction but can also recover the noise model.

Track: full conference paper

Paper Type: methodological development