Light-RepViTSR: ultra-lightweight super-resolution for real-time photoacoustic endoscopy in tumor biopsy

Real-time in situ biopsy marks a paradigm shift in clinical oncology by enabling immediate intraprocedural pathological diagnosis during endoscopy. Photoacoustic endoscopy (PAE) is a pivotal technology in this field that uniquely visualizes tumor microvasculature and hypoxia through the synergistic fusion of optical contrast and ultrasonic resolution. However, PAE's inherent resolution-speed tradeoff in raster scanning induces severe motion artifacts from physiological activities (e.g., peristalsis and respiration), critically compromising diagnostic reliability. Although deep-learning-based super-resolution (SR) techniques show promise for photoacoustic microscopy, their clinical translation to PAE is hindered by excessive computational demands and insufficient real-time performance. To overcome this limitation, we propose Light-RepViTSR, an ultra-lightweight SR reconstruction network based on the RepViT architecture and specifically optimized for real-time PAE. Our approach integrates the representational capacity of RepViT's re-parameterizable convolutional blocks while eliminating non-essential components (e.g., squeeze-and-excitation layers) to maximize computational efficiency. Comprehensive evaluation on a multi-source dataset-including 19 previously unseen murine cerebrovascular images and 18 self-collected plant vein images-demonstrates the superiority of Light-RepViTSR. The network consistently outperforms conventional methods across scaling factors ( $\times 2$ , $\times 4$ , and $\times 8$ ) to achieve significant improvements in PSNR (up to $+1.41$  dB at $\times 8$ ) and SSIM (up to $+0.047$ at $\times 2$ ) while reducing model size by $>99\%$ and inference time by $>60\%$ versus SRResNet. This study establishes a pathway toward practical real-time high-resolution PAE, demonstrating significant potential for enhancing in situ tumor biopsy accuracy.