Fig. 1. ML-LR-PAM system. (a) ML-LR-PAM microscopy system with Airy beam excitation. (b) Phase pattern, Airy beam obtained from the focal plane of the objective lens, and Airy beam profile in deep propagation direction. (c) Design of two-dimensional sparse-sampling raster scanning and multitask SW-CycleGAN-based super-resolution reconstruction combined with Airy beam artifact removal. (d) Airy beam PSF at a=0.5, 1, and 2. FS, full-sampling scanning. SS, sparse-sampling scanning (see Visualization 1).

Fig. 2. SW-CycleGAN principle and deep neural network architectures. (a) Unsupervised training is performed with blood vessel and microsphere simulation data. Forward and reverse generator transformations GX and GY are trained concurrently with corresponding discriminators DX and DY, which progressively improve their ability to classify generated synthetic images from true input examples. (b) Details of SDAM-based generator GX (see Visualization 2).

Fig. 3. Comparison of SW-CycleGAN and other methods from mouse brain microvascular dataset [5]. (a) The first column (LR) represents 2× downsampling of low-resolution, Airy-beam-based source images with a=0.5, a=1, and a=2. The second to fifth columns represent the combined RL + FD Unet, RL + Res U-net, proposed multitask learning of SW-CycleGAN, and GT in simulation data. (b) PSNR mean for all methods on the test set. (c) Relative degradation in PSNR when introducing larger model mismatches (see Visualization 3).

Fig. 4. Comparison verification of the PAM vascular structure maps of Airy beam side-lobes artifact removal by SW-CycleGAN and RL methods [5]. (a) Comparison of the results of SW-CycleGAN and RL methods on simulated vascular data. (b), (c) Quantitative analysis profiles in the simulation blood vessel maps corresponding to the blue and green dotted lines in (a) (see Visualization 4).

Fig. 5. The results from real data collected using ML-RL-PAM in various complex environments, including leaf skeletons, mouse cerebral vasculature, and zebrafish adult pigmented stripes, are shown. (a)–(c) depict Airy-beam-based PAM imaging of leaf skeletons, mouse cerebral vasculature, and zebrafish adult pigmented stripes, respectively. The first two columns compare imaging results using Gaussian and Airy beams. The third to sixth columns show results for LR Airy, Airy plus RL + FD Unet, Airy plus SW-CycleGAN, and HR, Airy plus SW-CycleGAN. LR denotes low resolution, and HR denotes high resolution. (d) presents the quantitative analysis corresponding to the white dashed lines L1–L3 in (a)–(c) (see Visualization 5).

Fig. 6. Verification of the generalization of the proposed method using the physiopathological tumor microenvironment of mouse skin cancers. (a), (b) show the characterization of vascular structures and blood flow information in the melanoma and basal cell carcinoma models, respectively. The second through fifth columns display the corresponding results for the yellow boxes highlighted in the first column, including LR Airy (a=1), Airy plus RL + FU Unet, Airy beam plus SW-CycleGAN maps, as well as the ground truth (GT) from the ROIs. (c) presents the quantitative analysis corresponding to the white dotted lines L1–L4 in (a) and (b) (see Visualization 6).