刺激ラマン光音響顕微鏡における信号分解技術の比較研究

In this paper, we present a comparative study of signal decomposition techniques in multispectral photoacoustic microscopy (MS-PAM) systems that are based on the use of stimulated Raman spectroscopy (SRS). The signal decomposition algorithms leverage the inherent delay between the arrival of two photoacoustic (PA) signals, one from 532 nm and the other from 558 nm. We evaluate four signal decomposition methods: fast Fourier transform (FFT), least squares (LSQ), cross-correlation (XCorr), and a deep learning (DL) approach based on a convolutional neural network (CNN) coupled with an autoencoder architecture. These methods are tested against ground truth PA signals generated separately at 532 and 558 nm wavelengths from the mouse brain. Results show that the FFT method consistently outperforms the others in terms of accuracy, while LSQ, XCorr, and DL's performance exhibits greater variability, especially at longer time delays. Finally, the performance of oxygen saturation estimation is evaluated using the best method (FFT).