Dual-parameter photoacoustic spectral analysis and machine learning for noninvasive monitoring of osteosarcoma treatment response.

Early and accurate assessment of chemotherapy response is crucial for bone tumor management; however, while microstructural changes occur during early tumor development, conventional imaging modalities often struggle to capture this subtle functional alteration. In this study, we propose a quantitative photoacoustic (PA) method integrating frequency-domain spectral analysis and statistical modeling to monitor treatment response in osteosarcoma. Using a photoacoustic computed tomography (PACT) system, murine osteosarcoma tumors were longitudinally imaged at 660, 680, and 760 nm. To characterize therapy-induced changes in tissue acoustic scattering and absorber morphology, the PA spectral slope and Nakagami-m parameter were extracted from the radio-frequency (RF) signals. Their temporal evolution was quantitatively analyzed at baseline and multiple post-treatment intervals (at 3-day intervals). Machine learning classifiers were employed to evaluate the discriminative capability of individual features and a multi-parameter fusion model. Results revealed distinct, treatment-dependent trajectories in both the spectral and statistical parameters, reflecting the evolution of the tumor microenvironment in terms of microstructural and functional alterations. While single-parameter models yielded Area Under the ROC Curve (AUC) values between 0.862 and 0.913, the fusion model significantly enhanced performance to an AUC of 0.950. These findings provide a preclinical feasibility that combining frequency-spectrum and statistical distribution features of PA signals offers a sensitive, label-free method for monitoring therapy-induced changes, highlighting the potential of quantitative PA signal characterization for future translational research in orthopedic oncology.