Title: A Noise-Aware Attention Mechanism for Enhancing Diffusion Weighted Imaging

Abstract:

While Diffusion-Weighted Imaging (DWI) is a valuable tool for comprehensive cancer screening, its utility is often constrained by lengthy acquisition times. Accelerating the scan process frequently compromises image fidelity, resulting in heightened noise levels. Specifically, the magnitude reconstruction inherent to DWI generates Rician noise that varies with the signal intensity. This signal-dependent characteristic poses significant challenges for traditional convolution-based denoising techniques.

To overcome these obstacles, we introduce a novel denoising framework driven by attention mechanisms and aware of noise characteristics, designed specifically for DWI restoration. This approach combines hierarchical window attention from Swin Transformers with a transformer-based multi-dimensional gated refinement module. By embedding explicit noise-level conditioning and utilizing residual reconstruction, the model adaptively suppresses heteroscedastic noise across diverse levels of image corruption.

Evaluations conducted on corrupted DWI scans highlight the framework’s superior restoration capabilities. Across noise levels ranging from 1% to 15%, the proposed model attained a mean Peak Signal-to-Noise Ratio (PSNR) of 33.69 dB and a Structural Similarity Index (SSIM) of 0.8539, while demonstrating consistent stability even under severe noise conditions. These findings suggest that integrating attention-guided contextual modeling with channel-adaptive refinement offers a robust and broadly applicable solution for denoising DWI data.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC