Structure and oxygen saturation recovery of sparse photoacoustic microscopy images by deep learning

Abstract

Photoacoustic microscopy (PAM) leverages the photoacoustic effect to provide high-resolution structural and functional imaging. However, achieving high-speed imaging with high spatial resolution remains challenging. To address this, undersampling and deep learning have emerged as common techniques to enhance imaging speed. Yet, existing methods rarely achieve effective recovery of functional images. In this study, we propose Mask-enhanced U-net (MeU-net) for recovering sparsely sampled PAM structural and functional images. The model utilizes dual-channel input, processing photoacoustic data from 532 nm and 558 nm wavelengths. Additionally, we introduce an adaptive vascular attention mask module that focuses on vascular information recovery and design a vessel-specific loss function to enhance restoration accuracy. We simulate data from mouse brain and ear imaging under various levels of sparsity (4 ×, 8 ×, 12 ×) and conduct extensive experiments. The results demonstrate that MeU-net significantly outperforms traditional interpolation methods and other representative models in structural information and oxygen saturation recovery.