[论文解读] Optical Inversion and Spectral Unmixing of Spectroscopic Photoacoustic Images with Physics-Informed Neural Networks

Sarkis Ter Martirosyan, Huang, Xinyue+3 more|arXiv (Cornell University)|Feb 18, 2026

Photoacoustic and Ultrasonic Imaging被引用 0

一句话总结

SPOI-AE 训练一个物理信息自编码器，在无 ground-truth 标签的情况下执行活体光声光谱图像的光学反演和光谱混合，优于传统线性方法。

ABSTRACT

Accurate estimation of the relative concentrations of chromophores in a spectroscopic photoacoustic (sPA) image can reveal immense structural, functional, and molecular information about physiological processes. However, due to nonlinearities and ill-posedness inherent to sPA imaging, concentration estimation is intractable. The Spectroscopic Photoacoustic Optical Inversion Autoencoder (SPOI-AE) aims to address the sPA optical inversion and spectral unmixing problems without assuming linearity. Herein, SPOI-AE was trained and tested on extit{in vivo} mouse lymph node sPA images with unknown ground truth chromophore concentrations. SPOI-AE better reconstructs input sPA pixels than conventional algorithms while providing biologically coherent estimates for optical parameters, chromophore concentrations, and the percent oxygen saturation of tissue. SPOI-AE's unmixing accuracy was validated using a simulated mouse lymph node phantom ground truth.

研究动机与目标

• Motivate accurate estimation of chromophore concentrations and tissue SO2 from spectroscopic photoacoustic (sPA) images.
• Address nonlinearities and ill-posedness in sPA optical inversion through a self-supervised deep learning framework.
• Develop a physics-informed autoencoder to jointly estimate optical parameters and chromophore concentrations.
• Demonstrate improved reconstruction quality and biologically coherent parameter estimates in in vivo mouse lymph node data.

提出的方法

• Model the sPA forward problem with diffusion-based optical transport and linkage to absorption spectra.
• Use SPOI-AE to separately estimate the absorption coefficient mu_a and reduced scattering coefficient mu_s' via two FCNNs (mu_a-Net and mu_s' -Net).
• Unmix chromophores by decomposing mu_a into weighted absorption spectra to obtain relative concentrations c.
• Employ a deterministic decoding stage that reconstructs p from estimated mu_a and mu_s' through a physics-based forward model (and a low-rank mu_a reconstruction).
• Train in a self-supervised manner using a combined loss that blends MSE and a wavelength-dependent spectral-angular-distance (MSAD) term.
• Compare against Lit. NLS and NMF baselines and evaluate with R^2 GoF across wavelengths.

实验结果

研究问题

• RQ1 Can a physics-informed autoencoder accurately invert the sPA problem and unmix chromophore spectra without ground-truth concentrations?
• RQ2 Does incorporating fluence compensation and optical transport physics improve spectral unmixing and SO2 estimation in vivo?
• RQ3 How does SPOI-AE performance compare to traditional linear unmixing (NLS, NMF) in terms of reconstruction error and wavelength-dependent GoF?
• RQ4 Can the absorption spectra be refined during training to better account for spectral coloring and nonlinear effects?

主要发现

Algorithm	MSE	MSAD
Lit NLS	0.0270	0.194
NMF	0.0259	0.198
Adjusted E, beta=5 (SPOI-AE)	0.0114	0.096
Adjusted E, beta=0 (SPOI-AE)	0.0124	0.116
Fixed E, beta=5 (SPOI-AE)	0.0126	0.100
Fixed E, beta=0 (SPOI-AE)	0.0140	0.122

• SPOI-AE variants with adjusted spectra and beta=5 achieve the best MSE (0.0114) and MSAD (0.096) on the testing set.
• Across wavelengths, SPOI-AE with adjusted E and beta=5 attains the highest R^2 GoF (0.897) with low variance (0.0331) on the testing set.
• SPOI-AE outperforms Lit. NLS and NMF in both MSE/MSAD and wavelength-wise GoF, demonstrating superior reconstruction of input sPA pixels.
• Refining absorption spectra during training improves spectral unmixing performance and GoF, indicating compensation for residual nonlinearities.
• The approach yields biologically coherent estimates for mu_a, mu_s', chromophore concentrations, and SO2 in vivo.