Multi-wavelength photobleaching kinetics of endogenous fluorophores for label-free tissue characterization with potential for tumor margin estimation using machine learning.

Autofluorescence photobleaching encodes valuable information related to tissue morphology, metabolism, and fluorophore microenvironment; however, it remains mostly underexplored as a quantitative imaging biomarker. This study investigates multi-wavelength photobleaching kinetics of endogenous fluorophores, including NADH, FAD, lipofuscin, and protoporphyrin, as label-free contrast biomarkers for breast tissue characterization with potential for tumor margin estimation. Time-lapse fluorescence images are acquired using a spectral imaging system equipped with a liquid crystal tunable filter under 405 nm excitation, selectively accessing fluorophore-specific emission wavebands. Instead of conventional double-exponential curve fitting, machine learning (ML) and neural network frameworks are employed to extract nonlinear relationships between photobleaching profiles and tissue pathology for pixel-wise segmentation. Furthermore, a physics-guided convolutional neural network (CNN), termed Exponential Feature Extraction Network (ExpFENet), is introduced, which incorporates prior knowledge of double-exponential photobleaching decay to guide the learning process for extracting discriminative features consistent with the underlying photobleaching behavior. The results showcase the importance of monitoring multi-waveband photobleaching dynamics in generating coherent tissue segmentation maps. Compared to ML and basic CNN architecture, ExpFENet achieves superior segmentation performance while improving robustness and interpretability. Overall, this study highlights that multi-wavelength photobleaching kinetics can serve as a promising label-free biomarker for tissue segmentation, with potential application to tumor margin estimation.