Using the latent diffusion model to enhance time-resolved laser speckle contrast imaging (TR-LSCI) of cerebral blood flow

Continuous monitoring of cerebral blood flow (CBF) with high spatiotemporal resolution and depth sensitivity is essential for accurate diagnosis and effective management of neurological disorders. Although conventional laser speckle contrast imaging (LSCI) enables widefield, high-resolution CBF mapping, its limited penetration depth and signal integration across all tissue layers hinder depth-resolved imaging. To address these limitations, we developed an advanced time-resolved LSCI (TR-LSCI) system that employs picosecond-pulsed laser illumination and a customized SPAD512² camera operating in gated mode, enabling noncontact, widefield, and depth-sensitive CBF imaging. However, photon scattering and diffusive noise still degrade image quality, particularly at greater depths. To overcome this challenge, we incorporated a multiscale latent diffusion model (LTDiff++) into the TR-LSCI analysis pipeline to suppress photon diffusion noise. Trained and validated using overlapping image patches from head-simulating phantoms and neonatal rat CBF images with high-quality ground truth references, LTDiff++ effectively suppressed photon diffusion noise while preserving structural and vascular features at greater imaging depths. Moreover, in vivo studies demonstrated that LTDiff++ maintained image quality using only 5-frame averaging, reducing acquisition time by a factor of 20 compared to the conventional 100-frame averaging approach without deep learning enhancement. The integrated TR-LSCI and LTDiff++ framework thus enables robust, high-speed, and depth-resolved imaging of cerebral hemodynamics, offering a promising platform for preclinical research and future clinical applications in bedside neuroimaging and patient monitoring.