Anisotropic 3D MRI Super-Resolution via Cross-Slice Collaborative Attention Propagation Network

In reference-based super-resolution reconstruction, feature transfer poses a significant challenge. For 3D magnetic resonance imaging, the structure can be viewed as a stack of successive slices, where adjacent slices typically exhibit structural similarity and spatial continuity. This inherent characteristic provides a feasible approach for leveraging information from neighboring slices to assist in the reconstruction of a target slice. We propose a cross-slice collaborative attention mechanism to exploit such continuity by adaptively transferring features from adjacent slices. The proposed attention mechanism employs a multi-granularity feature alignment strategy that reduces computational complexity by constraining the search space. Additionally, the model reuses the initial features extracted from the low-resolution input slice and then integrates them with high-information-content features from the reference slices through a residual fusion strategy, which enhances reconstruction quality. Experimental results on two public brain MRI datasets demonstrate that, under a magnification factor of ×2, our method improves the PSNR by 0.29 dB and 0.41 dB, respectively.