Abstract
Precise intraoperative delineation of tumor margin is critical for maximizing resection completeness and minimizing recurrence in head and neck cancers (HNC). Label-free Fluorescence Lifetime Imaging (FLIm), which captures the fluorescence decay characteristics of endogenous molecules, offers a real-time method for differentiating malignant from healthy tissue during surgery without the need of contrast agents. Here, we present a data-centric artificial intelligence (AI) framework to enhance the robustness and accuracy of FLIm-based classification models for HNC. FLIm data were collected in vivo from 92 patients undergoing both transoral robotic surgery (TORS) and non-TORS procedures for HNC using a multispectral FLIm device with 355 nm excitation. To improve model performance, a data-centric approach leveraging confident learning was implemented to identify and exclude instances with low quality. An interpretability framework was further integrated to quantify feature contributions and elucidate FLIm-derived sources of contrast. Under a leave-one-patient-out cross-validation scheme, the model demonstrated a strong discriminative ability with an area under the receiver operating characteristic curve of 0.94 in differentiating healthy versus cancerous tissue. In tumor boundary regions, borderline predictions revealed transitional tissue properties, with strong correlations observed between model predictions and FLIm parameters in spectral channels corresponding to NADH and FAD-key metabolic cofactors indicative of cellular metabolic shifts at tumor margins. The impact of tumor anatomical site (base of tongue, palatine tonsil, oral tongue) and p16+ (HPV) status on classification performance was also assessed. These findings underscore the potential of label-free FLIm to provide accurate, real-time guidance for intraoperative margin assessment and dysplasia grading, advancing surgical precision in head and neck surgical oncology.