Joint Multi-anatomy Training of a Variational Network for Reconstruction of Accelerated Magnetic Resonance Image Acquisitions

摘要

Magnetic resonance imaging is a leading image modality for many clinical applications; however, a significant drawback is the lengthy data acquisition. This motivates the development of methods for reconstruction of sparsely sampled image data. One such technique is the Variational Network (VN), a machine learning method that generalizes traditional iterative reconstruction techniques, learning the regularization term from large amounts of image data. Previously, with the VN technique, reconstruction of 4-fold accelerated knee images was shown to be highly successful. In this work we extend the VN approach to applications beyond knee imaging and evaluate the classic VN and a newly developed Unet-VN in 5 different anatomical regions. We evaluate the networks trained individually for each anatomical area as well as jointly trained with data from all anatomical areas. The VN and Unet-VN were trained to reconstruct 4-fold accelerated images of knees, brains, hips, ankles and shoulders. SSIM was calculated to quantitatively evaluate the reconstructed images. Results show that the Unet-VN outperforms the classic VN, both quantitatively - in terms of structural similarity - and qualitatively. The networks jointly trained with multi-anatomy data approach the performance of the individually trained networks and offer the simplicity of a single network for a range of clinical applications which has substantial benefit for clinical translation.