Diffusion Model‐Enhanced In Vivo Fluorescence Imaging for Spatiotemporal Vascular Visualization

Abstract

Near‐infrared (NIR) fluorescence imaging enables real‐time vascular visualization, but conventional indocyanine green (ICG) imaging is limited by shallow tissue penetration, low contrast, and poor spatial resolution. Lead sulfide (PbS) quantum dots in the NIR‐IIb window (1 500–1 700 nm) yield superior images yet face toxicity and regulatory barriers. Here, we introduce a diffusion model‐based fluorescence imaging framework that transforms raw ICG images into high‐resolution, high‐contrast outputs comparable to PbS imaging while retaining the safety and accessibility of clinically approved dyes. Using paired ICG and PbS vascular datasets, the model learns cross‐domain representations for high‐fidelity image‐to‐image translation. Our method preserves fine microvascular structures, including choke zones and fine branches, under low‐signal conditions. Quantitative evaluation revealed a nearly threefold increase in contrast, a 2.2‐fold enhancement in spatial resolution, and a signal‐to‐noise ratio improvement from 6.8 to 20.2 across a 6 × 5 cm field of view, while maintaining structural fidelity. Applied to perforator flap mouse models, the framework enabled spatiotemporal visualization of angiogenesis and vascular remodeling over 0–14 days, providing insights into vascularization in vivo. These findings demonstrate that integrating deep learning with fluorescence imaging platforms can overcome intrinsic optical limitations and offer a non‐invasive, spatiotemporal, and cost‐efficient alternative to expensive imaging modalities.