Results

We tested the ability of PaDI to resolve decorrelation speed maps that vary as a function of space and at different phantom tissue depths. PaDI reconstructions for two variable-speed perturbations under both 5 and 8 mm of turbid medium are in Figure 5C. In this experiment, 1280 and 108 patterns of variable speed and shape were utilized for training and testing, respectively. First, we observe that PaDI can spatially resolve features while still maintaining an accurate measure of unique decorrelation speeds. When structures with different decorrelation speeds begin to spatially overlap, the associated reconstructed speed values close to the overlap boundary are either lifted or lowered toward that of the neighboring structure. This is expected, as the detected light travelling through the “banana-shaped” light path contains information integrated over a finite-sized sensitivity region that will effectively limit the spatial resolution of the associated speed map reconstruction. Moreover, we also observed that PaDI reconstructions of dynamics hidden beneath a thicker 8 mm scattering medium are less accurate than those for dynamics beneath a 5 mm scattering medium. A resolution analysis based on the contrast in the edge regions of the circles, and the width of 10–90% edge response is also provided in Figure 5E. Speckle fluctuations sampled by our current configuration on the phantom tissue surface are less sensitive to decorrelation events occurring within deeper region. Creating a PaDI probe with larger source-detector separations can help address this challenge, as detailed in the Section 2.4. Further, we collect continuous data for 3 s, where the dynamic patterns hidden underneath present for 0.3 s, and change every 0.5 s. We show reconstructions of a few frames at Figure 5D. Figure 5F plots four of decorrelation rates change in time. The decorrelation rates are extracted by fitting each autocorrelation curves (using a Levenberg–Marquardt algorithm) with 1 + βeγτ, where τ is delay-time and γ is the decorrelation rates. These autocorrelation curves are used to generate reconstructions in Figure 5D. 3 s continuous measurements are taken, and the curves are estimated using 0.3 s integration window and 66.7% overlap between sliding windows. A set of autocorrelation curves from these four-fiber detection at 1.4 s is presented on the right.

We additionally conducted an experiment to study how our model, trained with data generated on our digital phantom, can reconstruct images of the dynamic scattering introduced by more biologically realistic contrast mechanisms. Noninvasive imaging of deep blood flow dynamics, such as hemodynamics within the human brain, is an important application for diffuse optical correlation-based measurements. Accordingly, we modeled deep hemodynamic flow by placing two capillary tubes (3 mm diameter) directly beneath a dynamic scattering volume (same optical properties: μa = 0.01 mm−1,) flowing at two different speeds (2.7 and 8.0 mm s−1) via syringe pump injection. After training an image formation model with PaDI data captured on our DMD-based phantom (630 maps of randomly oriented tube-like objects varying at 4–12 kHz), we acquired PaDI data from this unique capillary flow phantom and applied the DMD phantom-trained model to produce images as shown in Figure 6B. Here, we observe reconstructed image measurements of relative flow speed with spatial and temporal structures that match ground truth, pointing toward a system that can potentially image dynamic scattering beneath tissue in vivo using learning-based reconstruction methods trained with more easily accessible synthetic data.