Early detection of oral cancer using machine-learning enhanced multispectral autofluorescence lifetime endoscopy

In 2025, approximately 59,660 new cases of oral cavity and pharyngeal cancer are expected in the U.S. Although early detection significantly improves survival and quality of life, diagnosing oral cancer at an early stage remains challenging. Subtle symptoms, similarity to benign lesions, and limitations of visual examination contribute to delayed diagnosis. Despite the oral cavity’s accessibility, only 26% of cases are detected early, highlighting the need for affordable, accurate, and user-friendly diagnostic tools suitable for primary care and dental settings. Existing adjunctive methods—such as exfoliative cytology, vital staining, salivary diagnostics, and optical interrogation—have not demonstrated sufficient sensitivity or specificity. Intensity-based autofluorescence imaging, the most studied optical technique, suffers from low specificity due to signal loss in both benign and malignant tissues. In contrast, malignant transformation alters metabolic activity, producing distinct changes in autofluorescence lifetime and spectra. We have developed a cost-effective, dual-excitation, multispectral FLIM endoscopy system optimized for NADH and FAD sensing. Calibrated systems were deployed across four clinical sites, collectively imaging over 300 patients with oral lesions. Preliminary analysis using classical machine-learning (ML) models, trained and cross-validated on this dataset, shows promising diagnostic performance in distinguishing malignant from benign lesions (sensitivity: 0.83, specificity: 0.81, negative predictive value (NPV): 0.98). These findings underscore the potential of ML-enhanced endogenous FLIM endoscopy for reliable early oral cancer detection.