ADR-Net: Selective enhancement of photoacoustic images by deep learning

Photoacoustic microscopy (PAM) combines optical and ultrasound technologies to generate high-resolution images of biological tissues, enabling detailed visualization of microvascular structures. However, PAM imaging suffers from quality degradation in out-of-focused regions, which complicates data interpretation. Although recent methods focus on reconstructing the entire image uniformly, limitations remain in addressing selective enhancement of only the detected degraded regions. To overcome this challenge, this study proposes the anomaly detection and reconstruction network (ADR-Net), a deep learning framework designed to detect and enhance out-of-focused areas in PAM images, ensuring high-fidelity reconstructions. ADR-Net integrates a U-Net anomaly detector with a fully dense U-Net (FD U-Net) reconstruction module. The anomaly detection module generates an anomaly mask, guiding the reconstruction module to enhance degraded regions selectively. ADR-Net consistently outperforms baseline models including U-Net, FD U-Net, super-resolution diffusion (SRDiff), enhanced super-resolution generative adversarial networks (ESRGAN), and swin transformer-based image restoration (SwinIR) across key metrics, including SSIM, MSE, and PSNR. Our model achieves a PSNR improvement of approximately 19.87%, a 13.32% higher SSIM, and a 68.18% reduction in MSE compared to FD U-Net, along with a 70.0% lower anomaly detection (AD) score. This advancement demonstrates ADR-Net's potential for reliable and interpretable reconstruction of PAM images through selective enhancement of degraded regions, supporting both clinical and research applications.