hide
Free keywords:
-
Abstract:
Structural MRI is often prone to motion artifacts which are particularly pronounced in ultra-high-field MRI, leading to severe image degradation and negatively impacting downstream analysis. In this work, we propose a novel image similarity function that utilizes deep 3D features of pre-trained CNNs as a proxy for human perception of image similarity. We show that it outperforms classical metrics like L1, L2, PSNR and SSIM, and when used as a loss function it improves the performance of MRI motion correction models and address the limitations of per-pixel similarity measures commonly used in regression problems, which are inadequate in capturing structural and content-based differences between images.
