Integration of Deep Learning and Machine Learning for Label-Free Two-Photon Imaging-Based Classification of Oral Lesions: A Pilot Study

Early and objective diagnosis of oral squamous cell carcinoma (OSCC) and oral potentially malignant disorders (OPMDs) remains challenging due to limitations in conventional histopathology, including sampling bias, staining variability, and subjective interpretation. Label-free multiphoton imaging, particularly two-photon fluorescence (TPF), offers intrinsic contrast by providing endogenous signals from stromal autofluorescence (collagen, elastin, and basement membrane components) along with collagen-related second-harmonic generation, eliminating the need for exogenous dyes. In this study, we propose an integrated pipeline combining TPF imaging with ConvNeXt-Small network-based feature extraction and machine learning classification. Formalin-fixed, paraffin-embedded tissue sections from pathologist-confirmed cases were imaged using a custom two-photon system. Features from the penultimate layer of a ConvNeXt-Small convolutional neural network were extracted and classified using a multiclass support vector machine with radial basis function kernel. The model achieved an overall accuracy of 90%, with macro-average precision, recall, and F1-scores of 0.88, 0.91, and 0.89, respectively. Class-wise F1-scores were 0.99 for inflammatory/control, 0.78 for OPMD, and 0.91 for malignant. These findings demonstrate the potential of combining label-free TPF imaging with modern deep learning for rapid, objective, and label-free screening of oral lesions.