Ringing artefact removal from sparse view tomosynthesis using deep neural networks

Tomosynthesis imaging is a special type of X-raybased image acquisition wherein several projections are acquired from a limited angular range around the object or in a line above it. Subsequently, from these projection images, cross-section images parallel to the detector are reconstructed. However, if only a small number of projection images is acquired on a limited angular range, the reconstruction images suffer from ringing artefacts. This issue has not been addressed in the literature due to the lack of any model-based approach explaining this phenomenon. In the clinical setting, these artefacts can hinder a correct diagnosis. In this work, a deep learning-based ringing artifact reduction algorithm is proposed. The deep learning network was trained on 45786 medical images, resulting in a substantial reduction of ringing artefacts in the tomosynthesis reconstructions. Based on the numerical and visual evaluations, a conclusion is made on the positive effect of a deblurring Deep Neural Network in getting higher quality outputs.

Jaar van publicatie:2020