Large-volume photoacoustic microscopy has garnered significant attention within the biomedical photonics research community. This is due to its capability to visualize the intricate optical absorption properties of biological tissue with both high contrast and resolution [1–3]. In previous work, we introduced a flexible depth-of-focus (DOF), depth-invariant resolution photoacoustic microscopy (FDIR-PAM), employing non-diffraction Airy beams to facilitate large-volume imaging while upholding exceptional resolution [4]. Despite the evident superiority of FDIR-PAM, two inherent limitations persist, constraining its widespread adoption as a favored imaging modality.

The initial challenge confronting FDIR-PAM imaging pertains to the trade-off between spatial resolution and imaging speed [5]. Advancing clinical translation necessitates the imperative of rapid screening and repeatable monitoring in optical biopsy for clinical disease evaluation. While the FDIR-PAM imaging technique showcases notable advantages in delineating intricate histopathological biomarkers, the pace of imaging across expansive fields of interest remains sluggish due to the reliance on high-magnification objective lenses and traditional full-sampling point-by-point mechanical scanning [5]. This is particularly evident without compromising spatial resolution and the demand for large-volume imaging. Rapid optical biopsy and repeatable monitoring serve as pivotal measures for guiding PAM-assisted ablation and assessing treatment efficacy [6–8]. Conforming to Nyquist sampling theory, ensuring high spatial resolution in PAM images hinges on maintaining a full sampling scan step size typically not exceeding half of the anticipated spatial resolution [5]. Besides sophisticated and costly hardware methods [9–12], there is a growing preference for straightforward and economical software solutions that streamline system complexity. In modern high-speed microscopy imaging, a favored approach for enhancing imaging speed involves reducing the number of sampled points through sparse sampling [13]. Consequently, sparse sampling presents an indispensable compromise to improve imaging speed. The task of single-image super-resolution endeavors to reconstruct high-resolution images from sparsely sampled data, offering a significant improvement in imaging speed while preserving detection accuracy. In recent years, with the ascendancy of deep learning and deep image prior methods [14,15], image super-resolution reconstruction techniques tailored for under-sampled rapid PAM imaging have attained the pinnacle of performance without increasing system complexity. A secondary challenge that has plagued FDIR-PAM since it was proposed is the artifact-removed reconstruction task. In the realm of microscopic imaging, non-diffraction beams such as Airy beams have shown promise in effectively extending the DoF. However, despite the theoretical superiority of these non-diffracted beam imaging systems, practical implementation remains constrained. The primary hurdle arises from the side-lobes of non-diffraction beams, which induce a reduction in image contrast [16,17], particularly evident under low signal-to-noise ratios or in the presence of speckle and scattering, thus posing a severe challenge to traditional deconvolution algorithms [18]. Consequently, PAM systems employing non-diffraction beams often fail to yield superior contrast compared to Gaussian beams at the focal plane. In recent years, the advent of deep learning techniques has heralded new avenues to address this issue. Several studies have employed improved generative adversarial networks (GANs) to eliminate artifacts stemming from non-diffraction beam side-lobes in microscope images [19,20], with experimental findings demonstrating the superiority of deep-learning-based methods over traditional deconvolution approaches. While sparse reconstruction and artifact suppression tasks have been extensively investigated, they have conventionally been treated as distinct endeavors. In the realm of practical PAM for optical biopsy applications, images captured through non-diffraction Airy beam excitation and sparse scanning typically exhibit diminished spatial resolution and conspicuous side-lobe artifacts. Consequently, the synergistic execution of super-resolution reconstruction and side-lobe artifact removal emerges as paramount for achieving rapid, large-volume, and high-resolution PAM simultaneously.

In this study, we developed a hybrid hardware-software method that integrates Airy beam excitation and sparse sampling techniques to achieve rapid, large-volume imaging. This approach, termed multitask learning-powered large-volume, rapid photoacoustic microscopy with Airy beam excitation and sparse sampling (ML-LR-PAM), leverages the system’s known point spread function to simultaneously achieve super-resolution image reconstruction and remove Airy beam artifacts. These tasks are achieved through multitask learning, employing scaled dot-product attention mechanism (SDAM) Wasserstein-based CycleGAN (SW-CycleGAN). Experimental results confirm the efficacy of this method, facilitating the widespread adoption of Airy beam-sparse sampling combined with SW-CycleGAN in biomedical characterization. This modality offers rapid, large-volume imaging with high resolution and artifact-free performance. While existing techniques such as sparse sampling reconstruction and side-lobe suppression have demonstrated excellent results [5,21], our multitask learning approach provides a unique solution to the challenges of high-speed, extended DoF PAM, overcoming the trade-off between spatial resolution and imaging speed. Its key advantage lies in the integration of multiple tasks into a unified model, making it a powerful tool for large-volume, rapid imaging applications in fields like biological monitoring and functional imaging. This approach has the potential to complement and enhance existing methods, particularly in scenarios where both high speed and large DoF are essential. It is especially valuable for clinical applications, such as disease monitoring and treatment evaluation, where both spatial and temporal resolutions are critical for improving diagnostic accuracy and supporting real-time decision-making.

Figure 1(a) illustrates the configuration of the experimental setup. The setup architecture has been extensively described in our previous work, with recent minor adjustments made to certain optical path hardware details [4]. Here, we provide a concise outline. The Airy beam FDIR-PAM entails the utilization of a nanosecond pulsed laser (Poplar-532-10B, Huaray; wavelength 532 nm). A spatial light modulator (SLM, UPOLabs, HDSLM80R) operating in diffractive mode is employed to generate Airy beam illumination through the encoding of a flexible cubic phase pattern for precise phase modulation and laser intensity control. Subsequently, the modulated beam undergoes focusing and Fourier transformation by passing through the objective lens ($4\times$, 0.1 NA, UPlanSApo, Olympus) to obtain the desired Airy beam. The beam shaping module comprises a polarizer, half-wave plate, SLM, narrow beam lens group, objective lens, and three-dimensional moving platform. The polarizer and half-wave plate facilitate the polarization direction adjustment of the initial standard linearly polarized Gaussian beam, while the SLM performs phase modulation on the adjusted polarization to yield the modulated beam according to the pre-set azimuth phase modulation parameters. Notably, the SLM is also loaded with a cubic phase pattern featuring varying grating periods to not only produce the modulated light beam but also diffract it into distinct diffraction angles, facilitating the separation and subsequent filtering of the zeroth-order and first-order light spots using a pinhole. The cubic phase modulation parameters of the SLM, such as stripe spacing and modulation depth, can encompass multiple sets to modulate beams with diverse focusing lengths, main lobe sizes, and focused light energy. The Airy beam-sparse sampling PAM incorporates a ring-shaped ultrasonic transducer (UT) with a central opening (inner diameter: 1.2 mm; outer diameter: 4.5 mm; focal length: 4.4 mm; center frequency: 50 MHz; 6 dB bandwidth: 90%) for the detection of optically induced ultrasonic waves. Subsequently, the detected signals are amplified by two 30 dB low-noise amplifiers (ZFL-500LN+, Mini-Circuits), filtered through a 60 MHz low-pass filter (BLP 70+, Mini-Circuits), and acquired by a high-speed data acquisition board (DAQ, PCIE-1425, Yixing Technology) at a sampling rate of 250 MS/s. The SNR range in our current system can vary from 10 dB to 50 dB according to different imaging environments.

In the process of shaping the long focal length beam, the non-diffracting Airy beam is generated through phase modulation of a Gaussian beam using a phase mask and Fourier transformation with a Fourier lens. This technique enhances imaging resolution and DoF, providing significant advantages for biomedical imaging applications, particularly in capturing detailed structural and functional information. A high-resolution CMOS camera (DS-CFM300-H) was positioned at the rear focal plane of the objective lens to observe the evolution of the beam field distribution and the depth propagation direction of the non-diffused optical profile, as depicted in Fig. 1(b). The represented maps illustrate the phase pattern mask of the Airy beam, the sectional diagram captured at the focal plane of the objective lens, and the depth propagation direction of the Airy beam. To ensure image quality of high spatial resolution, full-sampling scanning adheres to the Nyquist sampling theorem, dictating that the scanning step size in PAM should be less than half of the lateral resolution of the system. Typically, the region of interest spans approximately $4\mathrm{mm}\times 4\mathrm{mm}$, with a scanning step of 2 μm, necessitating the collection of $2000\times 2000$ fully sampled raster scanning points. For the acquisition of fast sparse sampling data corresponding to the fully sampled data, under-sampled sparse scanning modes with scanning steps of 4 μm and 8 μm are employed, as illustrated in Fig. 1(c). In point raster scanning, imaging speed increases exponentially as the sparse sampling rate decreases. With a 10 μm scanning step, imaging speed increases approximately 25-fold compared to full-sampling of 2 μm step mode. Ultimately, the combination of Airy beams with sparse sampling of fast sparse data, along with restoration via a designed SW-CycleGAN reconstruction network, enables the attainment of a high-resolution source image with extended DoF. In Fig. 1(d), the Airy beam point spread function (PSF) images are displayed for different values of $a$ (0.5, 1, and 2), calculated and convolved with the experimental images to generate diverse datasets [18]. A dataset of microspheres or mouse blood vessel images produced through this method will be utilized for training and testing deep learning models. Specifically, a dataset comprising 500 images was randomly divided into a training set of 2,000 images and a test set of 50 images. The model training employed a learning rate of 0.002 and batch size of eight, and was conducted over 200 epochs on the laboratory platform.

The architecture of the SW-CycleGAN utilized in this study is illustrated in Fig. 2 and is derived from an enhanced CycleGAN model. Notably, CycleGAN offers a remarkable advantage due to its minimal requirement for paired training samples [22,23], substantially broadening its applicability. Specifically, CycleGAN streamlines data preparation by introducing two distinct sets of images, such as one containing an Airy beam image and another containing a Gaussian beam image without necessitating direct pairwise mapping. This approach opens up the potential for multiple image conversion tasks. A fundamental mechanism of CycleGAN is the cyclic consistency loss, which ensures the preservation of core content during image transformation. Consequently, the converted image not only maintains visual similarity to the original image but also achieves the desired shift in style or texture while preserving the image’s shape and structural features. This aspect is crucial for ensuring the quality of conversion, particularly when applied to microscopy image processing. These combined advantages render CycleGAN an ideal choice for Airy beam side-lobe removal, especially in scenarios where extensive data acquisition is unfeasible. It not only simplifies the complexity of data collection but also guarantees the quality and efficiency of the conversion process by preserving the fundamental structure and features of the image. However, CycleGAN is designed for the broader field of image processing and is not specifically tailored for the image super-resolution and Airy beam artifact removal required in this study. Therefore, an improvement was introduced based on CycleGAN to achieve the specified objectives, leading to the development of the improved network known as SW-CycleGAN. Specifically, a scaled dot-product attention mechanism (SDAM) [24] was integrated into the network, and WGAN’s W-loss [25] was utilized to optimize against GAN loss, as depicted in Fig. 2(b).

Self-attention mechanisms play a pivotal role in enabling models to discern and leverage intricate interrelations among different regions within an image. In the realm of vascular image processing, this capability empowers the model to concentrate on specific nuances and intricacies of the blood vessel, particularly those areas demanding meticulous attention during artifact removal and reconstruction. By enhancing the model’s comprehension of vascular architecture, especially amidst the presence of side-lobe artifacts, self-attention facilitates the model in discriminating between artifacts and authentic vascular structures, thereby facilitating more precise artifact removal and reconstruction. SDAM, a core calculation method of a typical self-attention mechanism in deep learning, holds significant utility in image processing domains. Its algorithmic flow, delineated in Fig. 2(b), involves the linear transformation of the input matrix X, represented as a series of “words”, into matrices Query (Q), Key (K), and Value (V). Herein, Q and K play pivotal roles in establishing semantic associations among the “words”, with the ensuing distance matrix reflecting these semantic linkages, computed through the core self-attention algorithm. The resultant output matrix, embodying global correlations, arises from the matrix multiplication of the distance matrix and matrix V. Wasserstein distance, also known as “bulldozer distance”, has showcased its efficacy across numerous studies since the inception of W-GAN [26,27]. Compared to conventional GAN loss, Wasserstein distance mitigates mode collapse during training by providing more continuous and useful gradients, which helps the generator learn a more accurate image transformation. When used in CycleGAN, Wasserstein distance enhances stability, promotes diversity in the generated images, and improves their overall quality. Additionally, it enables smoother and more realistic image transformations between domains. Hence, Wasserstein distance supplants the traditional GAN loss within CycleGAN to bolster the performance of the network proposed in this study.

To comprehensively assess the performance of the SW-CycleGAN network proposed herein, we introduce benchmark methods for comparison, namely, the sparse image recovery networks FD U-net and Res U-net, alongside the renowned Richardson-Lucy (RL) algorithm. Previous experiments have proven that, when dealing with sparse image restoration and Airy beam artifact removal, prioritizing artifact removal can improve image quality more effectively. Accordingly, in this study, we employ two control methods, namely, RL + FD Unet and RL + Res U-net, which first address Airy beam side-lobe artifact removal follow by super-resolution reconstruction. Simultaneously leveraging three Airy beam images with varying scale factor parameters ($a=0.5$, 1, and 2) and 2× downsampling rate to generate Airy beam, low-resolution (LR) images as inputs, we evaluate the performance of the SW-CycleGAN network. The results, depicted in Fig. 3(a), reveal that while the RL method diminishes side-lobes and aligns the image with the source, residual side-lobe remnants persist as background intensity. This residual presence compromises both resolution and contrast, resulting in an overall image quality reminiscent of lower resolutions. In contrast, the proposed SW-CycleGAN method demonstrates superiority in artifact removal as well as resolution improvement, rendering image details prominently visible and effectively reconstructing the subsampled Airy beam image, aligning it closely with the Gaussian source image in terms of quality. The mean peak signal-to-noise ratio (PSNR) was calculated across all 50 images utilized in the test data and is depicted in Fig. 3(b). Evidently, the SW-CycleGAN method outperforms the other two methods in terms of PSNR. Achieving PSNR values ranging from 20 to 23, the SW-CycleGAN method surpasses the combined RL + FD U-net and RL + Res U-net approach, both of which only attain PSNR values of 17 to 19. This underscores the significant enhancement in image quality facilitated by superior artifact removal effects. When employing traditional deconvolution techniques for image recovery in the presence of non-diffracting beam artifacts, precise system physical parameters are imperative for desired outcomes. Mismatches between the point spread function (PSF) utilized in deconvolution and the actual system PSF often yield subpar results. It has been substantiated that deep learning methods accommodate larger mismatches with greater efficacy [18]. In this study, the SW-CycleGAN method consistently demonstrates superior image quality, even in scenarios where discrepancies between test and training data are pronounced. Given that both resolution and artifact removal efficacy influence evaluation metrics, high-resolution images afflicted with side-lobe artifacts of Airy beams were employed to exclusively evaluate the SW-CycleGAN’s artifact removal performance in the presence of mismatched images alone. Figure 3(c) illustrates the relative degradation in PSNR when introducing larger mismatches relative to the source image. It is evident that the proposed SW-CycleGAN method tolerates larger mismatches compared to traditional RL methods, with an improvement in error measurement by approximately threefold.

Drawing on the unique effects of CycleGAN, SW-CycleGAN, trained for the two tasks of image super-resolution reconstruction and Airy beam side-lobe artifact removal, can also perform artifact removal independently. With this in mind, in order to further verify the artifact removal effect of SW-CycleGAN, a comparative experiment was conducted. First, the mouse blood vessel image with Gaussian beam imaging was generated into an Airy-beam-based PAM image ($a$ set to 1) by the simulation method mentioned above, and then artifact removal tasks were carried out by SW-CycleGAN and RL methods (Fig. 4). The arrows of different colors are used to mark three regions with different artifact influence degrees, among which the area pointed to the red arrow has a weak artifact influence. The blood vessel morphology can be observed in the recovery results of both SW-CycleGAN and RL methods, but obviously the result of SW-CycleGAN is better. In the region indicated by the blue arrow, the result of SW-CycleGAN recovery is exactly equivalent to that of a Gaussian-beam-based image, and all the vessel structure details are clearly distinguished, while the RL method can only roughly recover the structure of this region of interest. In the region marked by the green arrow, where vessel gaps are imperceptible in the Airy-beam-based image, SW-CycleGAN successfully restores the vessel gap structure, aligning closely with the Gaussian-beam-based image, surpassing the RL method’s performance. Hence, focused solely on the Airy beam artifact removal task, SW-CycleGAN clearly outperforms traditional deconvolution methods. Additionally, a quantitative analysis was conducted on the positions marked by the blue and green dashed lines in Fig. 4(a), where the blue dashed lines correspond to Fig. 4(b) and the green dashed lines correspond to Fig. 4(c). It is evident that SW-CycleGAN provides superior image quality compared to RL deconvolution, as demonstrated by the sharper curve where artifacts are effectively suppressed (green asterisks), closely resembling the results of the Gaussian-beam-based image.

To validate the effectiveness of the SW-CycleGAN in mitigating Airy beam side-lobes and enhancing image quality, particularly in comparison to conventional separate approaches, simulated vascular data were initially used. However, due to variations in system configurations, experimental settings, and inherent uncertainties, there may be discrepancies between simulated and real-world data. The performance of the proposed method has been further validated in various complex environments, including leaf skeletons, mouse cerebral vasculature, and zebrafish adult pigmented stripes [Figs. 5(a)–5(c)]. All data were acquired using an Airy-beam-based photoacoustic microscopy (PAM) system with a step size of $10\mathrm{\mu m}\times 10\mathrm{\mu m}$. For a $4\mathrm{mm}\times 4\mathrm{mm}$ region, the imaging time was approximately 13.3 min, representing a 25-fold increase in imaging speed compared to full-sampled $2\mathrm{\mu m}\times 2\mathrm{\mu m}$ step imaging. When comparing the imaging results of the Gaussian beam and the Airy beam (leftmost two columns), Airy-beam-based imaging shows a significant expansion of the DoF. The MAP (maximum amplitude projection) images provide more detailed information, and depth-direction profile maps successfully resolve deeper structures in both phantom and biological samples. However, the original Airy-beam-based MAP still exhibits side-lobe artifacts and background noise. To demonstrate the resolution enhancement and side-lobe artifact suppression capabilities of the proposed method, we first $2\times$ downsampled the Airy-beam-based images to generate low-resolution (LR) images (third column). This is equivalent to resulting in a 100-fold increase in imaging speed, reducing the imaging time for a $4\mathrm{mm}\times 4\mathrm{mm}$ region to approximately 3.3 min, compared to the full-sampled $2\mathrm{\mu m}\times 2\mathrm{\mu m}$ step. The comparisons between separate methods of Airy plus RL + FD Unet and Airy plus SW-CycleGAN are shown in the fourth and fifth columns of Fig. 5. The rightmost column shows the results of direct reconstruction of the Airy-beam-based images at a $10\mathrm{\mu m}\times 10\mathrm{\mu m}$ step size (HR, Airy plus SW-CycleGAN). Notably, the SW-CycleGAN method effectively removes background noise and suppresses Airy beam side-lobes, likely due to the style transfer effect of CycleGAN. At the same time, regarding vasculature reconstruction, the results of the conventional Airy plus RL + FD Unet method show distorted vascular morphology, while our proposed method produces smoother vascular structures with higher fidelity. Notably, the first derivative of the edge diffusion function at the zebrafish fringe boundary at the L3 position also exhibits improved resolution (as shown in the L3 illustration). Overall, the Airy plus SW-CycleGAN images demonstrate superior visual reconstruction, providing high-resolution, artifact-reduced images.

To further validate these results and highlight the impact of the improved resolution and signal-to-background ratio (SBR), we conducted a quantitative analysis [Fig. 5(d)] on the regions marked by the white dashed lines of L1–L3. The presence of a smaller peak, higher SBR, and improved resolution, as indicated by the blue and green lines in Fig. 5(d), demonstrates the superiority of the Airy plus SW-CycleGAN and HR, Airy plus SW-CycleGAN methods. These approaches effectively remove side-lobe artifacts and restore sparse images. Overall, these results clearly show that the SW-CycleGAN method significantly enhances the SBR and image resolution.

Finally, to further validate the generalization of the proposed method, its performance was evaluated and compared with high-quality co-registered Gaussian beam images obtained from the physiopathological tumor microenvironment of mouse skin cancers (Fig. 6), which depict vascular structures and blood flow information [7]. Airy beam PSF images with a scale factor value ($a=1$) were generated according to Ref. [18]. The first column illustrates the structural and functional characteristics of blood vessels associated with melanoma and basal cell carcinoma [Figs. 6(a) and 6(b)], with a step size of $10\mathrm{\mu m}\times 10\mathrm{\mu m}$. The second column shows $2\times$ downsampled low-resolution (LR) and Airy beam artifacts generated images from the yellow dashed boxes in the first column’s regions of interest (ROIs), representing low-resolution (LR), Airy-beam-based images. The third and fourth columns in Fig. 6 display the comparison of reconstruction results using two methods, Airy plus RL + FD Unet and Airy plus SW-CycleGAN. The results reveal that while the Airy plus RL + FD Unet method recovers contrast intensity information in some areas, the overall image quality is poor, with noticeable blurring and artifacts. In contrast, the proposed Airy plus SW-CycleGAN method preserves the structural and functional details of blood vessels with minimal image artifacts, closely matching the ground truth (GT) in the ROIs. Quantitative analysis across the ROIs (L1–L4) is shown in Fig. 6(c). The results clearly demonstrate that the Airy plus SW-CycleGAN method outperforms the Airy plus RL + FD Unet method, as evidenced by the sharper curve and lower noise, which indicates a more effective suppression of artifacts and superior image quality.

In this study, we developed a novel approach called multitask learning-powered large-volume, rapid photoacoustic microscopy with Airy beams (ML-LR-PAM), which utilizes SW-CycleGAN to drive Airy beam-sparse sampling photoacoustic microscopy (PAM). This method significantly advances the field of fast, high-volume imaging by combining hardware and deep learning techniques to achieve artifact removal and image super-resolution reconstruction. The Fourier transform of the Airy wave solution is given by a Gaussian beam with an additional cubic phase term [28]. To generate various Airy beam excitation, the pupil phase can be manipulated by imposing a cubic phase with different scale factors onto a Gaussian laser mode. This can be achieved, for example, using a spatial light modulator (SLM). Notably, SW-CycleGAN has been optimized by incorporating self-attention mechanisms and Wasserstein distances into the original CycleGAN network for effective image style migration. To validate the performance of SW-CycleGAN in removing Airy beam artifacts and improving image quality, extensive comparisons were made with traditional RL methods using simulated data sets. Remarkably, SW-CycleGAN outperformed the traditional methods even when physical parameters were unknown. Furthermore, the efficacy of SW-CycleGAN was verified on experimental images obtained from an Airy beam-sparse sampling PAM system. We anticipate that ML-LR-PAM will serve as a convenient and powerful tool for rapid, large-volume imaging and analysis, providing accurate image quality. Compared to traditional hardware methods, the multitask learning approach offers significant advantages in mitigating side-lobe effects. Conventional hardware solutions, such as spatial filtering, rely on fixed adjustments that may not adapt well to variations in imaging conditions, leading to suboptimal image quality. In contrast, the multitask learning method dynamically optimizes imaging parameters in real-time, effectively addressing side-lobe issues based on the specific imaging scenario. Additionally, this approach achieves super-resolution enhancement through software alone, eliminating the need for costly hardware upgrades and resolving the trade-off between image resolution and imaging speed. Overall, the multitask learning method provides greater flexibility, adaptability, and cost-effectiveness, making it suitable for dynamic environments and accessible to laboratories with limited resources. The most compelling application of the proposed method is in the clinical monitoring of brain diseases and atherosclerotic plaques. These conditions involve complex morphological changes that affect imaging sensitivity, particularly in deeper tissue layers. By enhancing image resolution and minimizing side-lobe artifacts, this advanced imaging technique enables high-quality, deep tissue imaging, improving diagnostic accuracy and monitoring disease progression. It allows for rapid, repeatable assessments, which are crucial for evaluating treatment efficacy and making timely clinical decisions, ultimately improving patient outcomes and supporting longitudinal studies. In the future, to achieve greater DoF and ultra-high resolution in pathological microscopic imaging with high numerical apertures, strategies such as Airy-beam-based multi-focus image fusion can be employed. This approach extends the DoF while maintaining focus across uneven tissue surfaces. Additionally, deep learning techniques like the SW-CycleGAN algorithm can enhance image quality and resolution, which is crucial for both clinical and high-throughput research applications. By using non-diffracting Airy beams, the DoF is extended without the need for axial refocusing, enabling faster and deeper imaging. The ML-LR-PAM technique enhances image resolution and reduces artifacts through efficient sparse data acquisition combined with the SW-CycleGAN algorithm. This approach overcomes the mechanical limitations and response time delays associated with traditional electrically tunable lenses (ETLs), allowing for high-speed imaging even at MHz laser repetition rates. Overall, the method improves imaging efficiency, depth capability, and image quality, offering greater versatility and compatibility across different imaging systems. Moreover, the superiority of ML-LR-PAM opens up possibilities for its integration with other imaging modalities, such as fluorescence microscopy, where real-time, large-volume feedback and high image quality are crucial.

In conclusion, ML-LR-PAM has been developed based on a combination of software and hardware solutions. The integration of Airy beam excitation and sparse sampling, coupled with SW-CycleGAN-based deep learning, facilitates rapid and high-volume imaging. By simultaneously addressing out-of-focus and under-sampled image resolution issues through this combined approach, we demonstrate improvements in image quality, including enhanced resolution and suppression of side-lobe artifacts. This new technique holds promise for applications in large-field volumetric imaging, rapid optical biopsy, as well as repeatable monitoring to guide PAM-assisted ablation procedures and assess treatment outcomes.