Multispectral photoacoustic tomography (PAT) offers high-resolution images of deep tissue oxygen saturation (sO2), but the complexity of photon absorption and scattering affects sO2 accuracy. This study applied a rigorous light transport model, revealing that PA spectra within biological tissue can be represented as convex cones (CCs) in a high-dimensional space. Using the CC model, sO2 can be estimated by finding the nearest CC to measured data, even in noisy conditions. This method combines a physical model with machine learning, demonstrating practicality and robustness in numerical, phantom, and in vivo imaging experiments, with an average sO2 estimation error of just 3% in human trials. Additionally, it outperforms clinical practices like linear spectral unmixing, suggesting broader applications in PA molecular imaging and diffuse optical imaging.
|