Physics-Informed Neural Network for Mapping Vascular and Tissue Dynamics Using Laser Speckle Contrast Imaging.

Area of Science:

• Biomedical Optics
• Neuroimaging
• Artificial Intelligence in Medicine

Background:

• Laser Speckle Contrast Imaging (LSCI) is crucial for studying cerebral blood flow, neural-vascular coupling, and stroke.
• Conventional LSCI analysis methods are slow and difficult to scale, hindering real-time applications.
• There is a need for efficient, physics-based methods to extract vascular and tissue dynamics from LSCI data.

Purpose of the Study:

• To develop and validate a physics-informed neural network (PINN) for quantitative estimation of vascular and tissue dynamics from LSCI.
• To achieve direct estimation of fast (vascular) and slow (tissue-related) speckle decorrelation parameters without ground-truth labels.
• To enable rapid, pixel-wise analysis of full-field LSCI measurements.

Main Methods:

• Developed a PINN integrating an analytical LSCI model into the network's loss function for physics consistency.
• Employed a self-supervised learning approach for pixel-wise inference across LSCI images.
• Validated the PINN framework using in vivo mouse stroke LSCI datasets.

Main Results:

• The PINN accurately recovered fast decorrelation rates (cerebral blood flow) and slow dynamics (tissue/cellular motion).
• Generated parameter maps comparable to traditional methods but achieved analysis speeds orders of magnitude faster (seconds vs. hours).
• Demonstrated generalization to unseen subjects and robustness under noisy conditions.

Conclusions:

• Physics-informed learning provides a practical framework for near real-time extraction of vascular and cellular biomarkers from LSCI.
• This method enables efficient longitudinal monitoring of stroke progression.
• The approach holds potential for facilitating clinical translation of LSCI-based diagnostics.