Virtual-point-based deconvolution for optical-resolution photoacoustic microscopy

Abstract

Optical-resolution photoacoustic microscopy (OR-PAM) has been increasingly utilized for in vivo imaging of biological tissues, offering structural, functional, and molecular information. In OR-PAM, it is often necessary to make a trade-off between imaging depth, lateral resolution, field of view, and imaging speed. To improve the lateral resolution without sacrificing other performance metrics, we developed a virtual-point-based deconvolution algorithm for OR-PAM (VP-PAM). VP-PAM has achieved a resolution improvement ranging from 43% to 62.5% on a single-line target. In addition, it has outperformed Richardson-Lucy deconvolution with 15 iterations in both structural similarity index and peak signal-to-noise ratio on an OR-PAM image of mouse brain vasculature. When applied to an in vivo glass frog image obtained by a deep-penetrating OR-PAM system with compromised lateral resolution, VP-PAM yielded enhanced resolution and contrast with better-resolved microvessels.

Keywords: deconvolution; genetic algorithm; photoacoustic microscopy.

FIGURE 1

Imaging principle of OR-PAM. (a) Schematic of a representative OR-PAM. OL, objective lens; UT, ultrasonic transducer. (b) Cross-sectional view of the focal region circled in (a). Absorbers within the optical focus contribute the most to the detected signal.

FIGURE 2

Optimization of VP-based deconvolution algorithms. The loss between the raw image $S$ and the estimated virtual image $\widehat{R}$ convolved with a known PSF is updated for each iteration. $\widehat{R}$ should resemble the ground truth if the loss is small enough.

FIGURE 3

Principle of implementing VP-PAM. (a) Schematic of the genetic algorithm implemented in VP-PAM. Step 0: Virtual points are randomly spread over a virtual image and form the initial children population. Step 1: Children with positive fitness scores are selected and become parents for the new generation, while the rest are discarded. Step 2: Low fitness parents are randomly discarded. The sifted parents survive and join the parent population of the next generation. Step 3: Sifted parents produce children in a mitosis-like manner—new children are randomly generated within the $3\times 3$ pixel Moore neighborhood. Higher fitness parents are likely to produce more children in each generation. Step 4: A small portion of children undergo mutation—small, random perturbations are applied to their genes. Step 5: Repeat Steps 1–4 until the maximum number of generations is reached. The final population consists of the sifted parents and the children of the last generation. (b) General workflow showing how the fitness score is updated and utilized. The population size stays the same.

FIGURE 4

Resolution enhancement on a synthetic single-line target by various methods. (a) Noise-free blurred single-line target. The blue line represents the ground truth. (b-f) Deblurred VP-PAM results from (a) with $\lambda =0$, 0.1, 1, 1.5, and 2, respectively. (g-h) Deblurred results from (a) using RL15 and A-PoD, respectively. (i) Blurred target with added noise. (j-n) Deblurred VP-PAM results from (i) with different $\lambda$ values. (o-p) Deblurred results from (i) using RL15 and VP-PAM, respectively. (q) Line profiles of (b)-(h) along the white line in (a). (r) Evolutions of mean squared error (MSE) of (b)-(f). Scale bar: 75 µm.

FIGURE 5

VP-PAM on a mouse brain vasculature image. (a) Ground truth image of mouse brain vasculature. (b) Blurred image. (c-e) Deblurred images using RL15, VP-PAM, and A-PoD, respectively. For better visualization, (e) was clipped between 0 and 1. (f-g) Line profiles of (a)-(e) along the two solid lines, respectively. Note that (e) was plotted using the right y-axis in (f) and (g). Scale bars: 1 mm for the main images and 400 µm for the insets.

FIGURE 6

VP-PAM on a glassfrog vasculature image. (a) Original D-PAM image of glassfrog vasculature. (b) Deblurred image by VP-PAM. (c-d) Close-up images of the boxed region #1 in (a), (b), respectively. (e-f) Close-up images of the boxed region #2 in (a), (b), respectively. (g) Line profiles of (c) and (d) along the solid line path in (c). (h) Line profiles of (e) and (f) along the solid line path in (e). Scale bars: 2 mm in (a-b), and 500 µm in (c-f).