Toward Clearer NIR-II Fluorescence Imaging: Unpaired Enhancement via Hierarchical Frequency Filtering and Spatial-Channel Contrastive Learning

Fluorescence imaging in the second near-infrared window (NIR-II) has great clinical potential for intraoperative visualization. NIR-IIb enables deeper tissue penetration than NIR-IIa. However, current NIR-IIb imaging probes have insufficient biosafety, NIR-IIa imaging remains the primary option in clinical practice. In this study, we propose a Hierarchical Frequency Filtering and Spatial-Channel Contrastive Learning Network (HF$^{2}$SC$^{2}$L-Net) to achieve unpaired image enhancement from NIR-IIa domain to NIR-IIb domain. HF$^{2}$SC$^{2}$L-Net contains three core modules: Hierarchical Frequency Filtering (HFF), Spatial-Channel Residual blocks (SSC-Res), and Query Selected Attention (QS-Attn). HFF decomposes NIR-IIa images into cross spectral invariant vascular structures and band specific spectral features, reducing spectral distortion. SSC-Res suppresses redundant regions and noninformative channels, enabling deeper feature extraction and model lightweighting. QS-Attn integrated with PatchNCE loss, focuses on high information regions for fine grained cross domain alignment. Experimental results demonstrate that HF$^{2}$SC$^{2}$L-Net generates high quality NIR-IIb images. Validated on public datasets, it achieves Fréchet Inception Distance (FID) of 37.77 on NIR-II mice imaging (NIR-II MI) and Peak Signal to Noise Ratio (PSNR) of 21.44 dB on NIR-II light sheet microscopy (NIR-II LSM), outperforming other competitive algorithms. Furthermore, our method obtains optimal FMT reconstruction results based on cross domain enhanced images from in-house dataset with Localization Error (LE) of 0.19 mm and Dice coefficient (Dice) of 0.73. The method enables improved tumor localization and vascular visualization, providing a potential technical foundation for future clinical translation.