Robust multi-contrast mri denoising using trainable bilateral filters without noise-free targets

Magnetic resonance imaging (MRI) acquisitions inherently suffer from noise which can be reduced using image denoising algorithms. Most MRI denoising methods are based on either non-linear filters or convolutional neural networks (CNNs). In this work we propose to combine the best of both worlds by employing a set of trainable bilateral filter (BF) layers without ground-truth data using an adapted version of Stein's unbiased risk estimator (SURE) to enable model-based supervision. Our quantitative and qualitative experiments suggest that a neural network consisting of trainable BF layers outperforms a standard CNN when tested on unseen MR image contrasts by 14.7% in terms of PSNR. Thus, the proposed method provides not only a parameter efficient but also a considerably more robust alternative to deep neural networks, making it particularly valuable for applications in clinical practice.