Physics-Informed Deep Biophotonics for Real-Time Cancer Tissue Characterization Under Optical Scattering Uncertainty

Accurate real-time cancer tissue characterization remains challenging because optical signatures are continuously affected by scattering variability, attenuation imbalance, and heterogeneous tissue microstructures. This study presents a physics-informed multispectral biophotonic imaging framework for robust cancer characterization under heterogeneous optical propagation conditions. The proposed framework integrates multispectral acquisition (450–850 nm), scattering-aware stabilization, propagation-consistent feature learning, lightweight physics-guided regularization, and uncertainty-aware inference within a unified architecture. A dataset comprising 12,480 multispectral tissue images obtained from 86 specimens collected from 43 patients was utilized, with histopathological assessment serving as the diagnostic ground truth. Patient-independent evaluation demonstrated an overall classification accuracy of 96.8%, sensitivity of 95.9%, specificity of 97.2%, F1-score of 96.0%, and AUC exceeding 0.96. The incorporation of propagation-aware constraints improved robustness against scattering-induced spectral degradation while preserving physically meaningful spectral–spatial representations. Uncertainty-aware inference successfully localized diagnostically ambiguous regions and reduced overconfident predictions under degraded optical conditions. Comparative analysis against CNN, ResNet, and transformer-based models confirmed superior robustness, interpretability, and computational efficiency. Furthermore, the framework maintained real-time performance with an average inference latency below 38 ms per frame. These findings demonstrate that integrating tissue–light interaction knowledge with artificial intelligence substantially improves reliability and translational applicability in computational biophotonics. The proposed framework provides a clinically interpretable and deployment-ready solution for real-time optical cancer tissue characterization and intraoperative decision support.