A novel channel-encoder latent diffusion models for in vivo human brain spectral reconstruction

Hyperspectral images (HSIs) provide neurosurgeons with a powerful tool to accurately identify and remove brain tumors while minimizing damage to healthy tissues. However, conventional HSI systems suffer from insufficient temporal resolution, inadequate spectral resolution for precise tissue differentiation, and high hardware costs that limit clinical accessibility. To address these challenges, we propose a spectral reconstruction method based on a latent diffusion model (LDM), termed the channel encoder latent diffusion model (CE-LDM), which reconstructs high-fidelity HSI from snapshot RGB images captured by standard cameras. This snapshot-based acquisition inherently resolves temporal constraints by freezing tissue motion, while dramatically reducing hardware costs through the use of conventional Red–Green–Blue (RGB) sensors. Given the high-dimensional nature of hyperspectral data, we design a channel encoder to extract prior information, reduce data redundancy, and effectively mitigate the underconstrained problem in spectral reconstruction. Meanwhile, the diffusion model significantly improves spectral fidelity through iterative denoising. Additionally, we design a residual attention denoising network (RADNet) to enhance feature extraction and image detail reconstruction. Tests on in vivo Human Brain HSI dataset validate demonstrates that CE-LDM achieves state-of-the-art reconstruction quality with the lowest computational cost, offering an optimal balance for real-time intraoperative applications, and subsequent classification tasks confirm its clinical applicability and great potential in advancing hyperspectral brain tumor diagnosis.