Deconvolution-enhanced self-supervised learning for photoacoustic microscopy denoising

Optical-resolution photoacoustic microscopy is an emerging imaging technique that enables high-resolution visualization of biological structures. However, the presence of system noise significantly degrades image quality. As a key post-processing step in photoacoustic imaging, denoising based on deep learning is limited by the small size of available training datasets, which restricts its application in high-precision imaging of in vivo tissues under complex noise conditions. To address the issue, we built a photoacoustic microscopy and proposed a deconvolution-enhanced self-supervised learning framework for denoising. The system integrates the network, enabling training on noisy measurements and operation with just a single photoacoustic image. The method generates photoacoustic microscopy training pairs using a directional neighborhood random down-sampling strategy and employs deconvolution to preserve vascular continuity and noise independence. A self-constrained loss function is designed to improve data utilization and improve the effectiveness of noise suppression. Experimental results on phantom, mouse ear, and mouse brain demonstrate that the proposed system and method effectively suppress intrinsic noise and reveal clearer structural details. Compared with the raw data, the average signal-to-noise ratio, contrast-to-noise ratio, and Brenner gradient index across these datasets increase by approximately 20%, 40%, and 30%, respectively. The proposed method can achieve denoising using only noisy photoacoustic data. The method and system demonstrate strong performance in high-fidelity photoacoustic imaging, showing great potential for biomedical applications.