[A vessel-enhanced consistency-preserving diffusion model for 3T-to-7T susceptibility weighted image generation]

To develop a method for generating 7T susceptibility weighted image (SWI) from 3T SWI data for more accurate vascular structure reconstruction.

We propose a vessel-enhanced consistency-preserving diffusion model for synthesizing 7T SWI. This method first utilizes a pre-trained vessel segmentation network to extract vessel maps from both 3T and 7T SWI as vascular priors. The vessel maps from 7T SWI are employed to construct an intra-slice vessel-enhanced loss and reinforce the reconstructed fine vascular structures. An inter-slice feature fusion module and a vessel-aware consistency constraint are integrated into the diffusion model to enhance the spatial coherence of the generated results. The feature fusion module integrates contextual features from adjacent slices via a cross-attention mechanism, enabling deep information interaction across multiple slices. The vessel-aware consistency constraint employs an inter-slice vessel-aware consistency-preserving loss to impose constraints on overlapping regions of the sliding windows, thereby enhancing generation coherence across slices.

The experiment conducted on 209 paired 3T-7T SWI datasets showed that, compared to the suboptimal method EDM, the proposed method increased the PSNR by 1.39 dB, improved the SSIM by 2.28%, enhanced the skeleton coverage similarity (SCS) by 6.57%, and reduced MAE and LPIPS by 15.84% and 17.07%, respectively. Qualitative analysis further validated the model's superiority in generating vascular details, including accurate restoration of major venous structures and clear delineation of low-contrast microvessels, with natural continuous vascular branches. Ablation studies confirmed the effectiveness of each individual module and their contribution to the performance gains.

The proposed method can effectively synthesize 7T SWI images with clear vascular details from 3T SWI to facilitate clinical and scientific analyses while balancing computational resource requirements and image reconstruction quality.