Abstract
Fiber laser sensors offer significant advantages for photoacoustic microscopy (PAM), including compact size, electromagnetic immunity, and suitability for fast scanning systems. However, its signal-to-noise ratio (SNR) may rapidly degrade when the field of view (FOV) is enlarged. This compromised SNR adversely affects the accuracy of blood oxygen saturation (sO(2)) derived from noisy photoacoustic signals. To address this problem, a two-stage deep learning framework for fiber laser sensor-based PAM is proposed. The first stage reduces the 3D data to 2D image and suppresses the noises. The second stage integrates the dual-wavelengths images and suppresses the spectral distortion, so that the accuracy of sO(2) can be preserved. The network performance is validated using imaging datasets acquired with a conventional high-SNR photoacoustic microscopy system. Results demonstrate that this approach does not only denoise images acquired with the unfocused fiber laser sensor, but also maintains high fidelity in sO(2) calculation, addressing a key challenge in fast functional PAM.