Controlling spatial correlation in k-space interpolation networks for MRI reconstruction: denoising versus apparent blurring
Controlling spatial correlation in k-space interpolation networks for MRI reconstruction: denoising versus apparent blurring
Purpose: To improve the interpretability of noise amplification and apparent blurring of k-space interpolation networks, and to optimize for them in the loss function as a model-based regularizer in k-space interpolation networks. Methods: Network is subjected to noise amplification analysis through automatic differentiation of the input with respect to the input. Noise variance maps are decomposed into terms accounting for the linear and nonlinear characteristics of the network. Variance maps are derived in each iteration, allowing for runtime quality monitoring. Maximum variance (eigenpixel) and residual variance maps (pixel contamination) are introduced, which describe the network noise amplification and apparent blurring, respectively. By including the variance maps in the training, the loss function is enriched with a model-based regularizer beyond the k-space data consistency term. Accordingly, the proposed g-factor-informed RAKI (GIF-RAKI) establishes a recurrent flow of noise and apparent blurring information into the training, that drives the denoising via the trainable nonlinear activation function. Results: GIF-RAKI outperforms other RAKI implementations, supported by difference maps, and image quality metrics. Eigenpixel and pixel contamination maps provide quantitative metrics for noise amplification and apparent blurring, respectively, without the need for a gold standard reference. RAKI with tuneable Leaky ReLU is capable of adjusting its own nonlinearity automatically. Conclusion: The additional model-based loss terms allow to optimize for the trade-off between denoising and apparent blurring during RAKI training. This has the potential to eliminate the need for heuristic hyperparameter tweaking.
Istvan Homolya、Peter Dawood、Jannik Stebani、Felix Breuer、Grit Hein、Matthias Gamer、Florian Knoll、Martin Blaimer
计算技术、计算机技术
Istvan Homolya,Peter Dawood,Jannik Stebani,Felix Breuer,Grit Hein,Matthias Gamer,Florian Knoll,Martin Blaimer.Controlling spatial correlation in k-space interpolation networks for MRI reconstruction: denoising versus apparent blurring[EB/OL].(2025-05-16)[2025-06-06].https://arxiv.org/abs/2505.11155.点此复制
评论