judgeMI: Towards Accurate Metrics for Assessing Deep Learning Based Structural MRI Motion Correction
Ultra-high field (UHF) MRI often suffers from motion artifacts, which can degrade image quality and affect subsequent analysis. In our study, we present an innovative solution to this problem by using deep features from pre-trained convolutional neural networks (CNNs) to correct motion artifacts in MRI.
Our approach includes a new image similarity function, that we call judgeMI, which mimics human perception of image similarity and has shown promising results in other applications. This function outperforms traditional metrics such as L1, L2, PSNR, and SSIM, especially when used as a loss function in MRI motion correction models.Traditional per-pixel measures fall short in capturing structural and content-based differences between images, especially in regression problems. To overcome this limitation, we employ a fully 3D convolutional VGG19 architecture in a loss network (denoted Φ) trained using a self-supervised learning strategy.
In this strategy, surrogate classes are created by applying different transformations (blur, noise, bias field, ghosting, anisotropy) to the original image. The network learns to discriminate between these surrogate classes by acquiring feature representations that are not task-specific, but more generic and versatile. This promotes the learning of meaningful features for judgeMI. Our evaluation includes two experiments. First, we synthetically generate weakly and strongly motion-corrupted images and compare the proposed similarity measure with classical metrics as a baseline. In the second experiment, we use a dataset of 100 9T brain MRI images to train a common MRI motion correction deep neural network and compare the proposed similarity measure with the standard L1 loss.
The results demonstrate the superiority of our proposed similarity function judgeMI, which achieves 68.6% accuracy in discriminating motion planes. The 3D UNet trained with our similarity function outperforms the traditional L1 loss, visibly improving image quality and preserving anatomical structures. Importantly, this improvement extends to unseen data, as indicated by lower L1 and L2 errors and higher PSNR and SSIM values on the test set. Our results suggest the potential of our similarity function to advance MRI motion correction networks beyond traditional loss functions.