Deep Learning Enables Superior Photoacoustic Imaging at Ultralow Laser Dosages

Abstract

Optical-resolution photoacoustic microscopy (OR-PAM) is an excellent modality for in vivo biomedical imaging as it noninvasively provides high-resolution morphologic and functional information without the need for exogenous contrast agents. However, the high excitation laser dosage, limited imaging speed, and imperfect image quality still hinder the use of OR-PAM in clinical applications. The laser dosage, imaging speed, and image quality are mutually restrained by each other, and thus far, no methods have been proposed to resolve this challenge. Here, a deep learning method called the multitask residual dense network is proposed to overcome this challenge. This method utilizes an innovative strategy of integrating multisupervised learning, dual-channel sample collection, and a reasonable weight distribution. The proposed deep learning method is combined with an application-targeted modified OR-PAM system. Superior images under ultralow laser dosage (32-fold reduced dosage) are obtained for the first time in this study. Using this new technique, a high-quality, high-speed OR-PAM system that meets clinical requirements is now conceivable.

Keywords: deep learning; multitask residual dense networks; optical‐resolution photoacoustic microscopy; ultralow laser dosage.

Figure 1

Schematic of the optical‐resolution photoacoustic microscopy (OR‐PAM) imaging system. CL: convex lens; FS: fiber beam splitter; FOA: fiber optic attenuator; RM: reflection mirror; OBJ: objective; EA: electronic amplifier; UST: ultrasonic transducer; PD: photodiode detector.

Figure 2

The overall framework of the proposed multitask residual dense network (MT‐RDN) method.

Figure 3

a–l) Quantitative analysis of peak signal to noise ratio (PSNR) of testing data: brain data are shown in the red dotted frame, and ear data are shown in the green dotted frame. The x‐coordinate represents the value of PSNR and the y‐coordinate represents the number of samples. Sample size = 720.

Figure 4

a–l) Quantitative analysis of structural similarity index (SSIM) of testing data: brain data are shown in the red dotted frame, and ear data are shown in the green dotted frame. The x‐coordinate represents the value of SSIM and the y‐coordinate represents the number of samples. Sample size = 720.

Figure 5

Comparison of the image quality before and after MT‐RDN. a) Input 1 of Data 2, b) Ground truth 1 of Data 2, c) Ground truth 3 of Data 2, d) Recon 3 of Data 2, scale bar = 0.5 mm.

Figure 6

Validation of the universal applicability of the MT‐RDN by showing the key endpoints of the network when using Data 5 as the input. a–c) Input 1, Ground truth 1, and Recon 1, respectively; d–f) Input 2, Ground truth 2, and Recon 2, respectively; g–i) filtered image of Inputs 1 by PAIVEF, Ground truth 3, and Recon 3, respectively; and j) quantitative analysis of SNR of the selected area, where the selected areas are indicated by the solid lines in (a3), (b3), (c3), and (i3). Scale bar = 0.5 mm.

Figure 7

a) U‐net reconstruction results of the ear imaging data. b) RDN reconstruction results of the ear imaging data. c) Recon 3 of the MT‐RDN reconstruction results of the ear imaging data. d–f) The error maps of (a)–(c), respectively, with Ground truth 1. Scale bar = 1 mm.

Figure 8

a–c) The sO2 maps use the original image, Frangi's filtered filter image, and Recon3 of MT‐RDN as the mask. Scale bar = 1 mm.