BIDeM-SR: blind isotropic MRI super-resolution via dual-input degradation modeling

Deep learning-based isotropic super-resolution (SR) methods for magnetic resonance (MR) imaging hold significant promise for reconstructing isotropic MR images from anisotropic inputs. However, most existing approaches primarily focus on in-plane degradation and reconstruction, often neglecting accurate modeling of through-plane degradation. We propose a blind isotropic SR method based on dual-input degradation modeling. Our approach employs a non-paired volumetric MR image dual-input architecture to capture the through-plane degradation process more effectively. This design addresses the limitations of conventional single-input frameworks, which rely on in-plane directional features and therefore fail to simulate actual through-plane degradation, ultimately degrading SR reconstruction quality. To mitigate feature representation bias when processing multi-scale MR images, we incorporate a scale-aware module based on conditional convolutions within the degradation network. This module dynamically adapts to scale variations, enhancing the network's ability to model complex degradations. Extensive experiments conducted on the ITKTubeTK and BraTS2013 datasets across multiple scales demonstrate that the proposed method outperforms existing single-input approaches in modeling through-plane degradation and achieves superior isotropic SR reconstruction performance.