D‑Blur: A Deep Learning Approach for Mapping Subdiffraction Diffusion with Motion-Blurred Images

Image-based single-particle tracking (SPT) provides insight into complex transport within diverse biological and porous material structures, but its performance is constrained by motion blur and a low signal-to-noise ratio (SNR). Traditional SPT methods are sensitive to localization errors and often struggle with short trajectories and fast-moving emitters. In this work, we develop D-Blur, a U-Net-based convolutional neural network (CNN) algorithm designed to localize single particles and predict their diffusion coefficients (D) from motion-blurred point spread functions (PSFs). The obtained D values of emitters enable the reconstruction of diffusion maps on confined transport in porous materials. We validate the algorithm with simulated emitters in a heterogeneous environment, as well as the experimental data of free diffusers in a controlled diffusion environment. By directly extracting molecular dynamics from microscopy images without requiring trajectory linking, D-Blur overcomes key limitations of conventional SPT, providing a solution for subdiffraction diffusion maps within the native imaging flow of fluorescence microscopy. This work enhances diffusion analysis in complex systems and lays the foundation for future applications.