Work

Combined feature-based classification of premalignant and malignant oral lesions using VELscope endogenous fluorescence imaging and dry serum Raman spectroscopy

Year: 2025

Type: article

Abstract: Accurate differentiation between premalignant and malignant oral lesions is critical for improving clinical management and patient outcomes. This study explores a multimodal diagnostic approach, uniquely combining endogenous fluorescence imaging, Visually Enhanced Lesion Scope (VELscope) and dry serum Raman spectroscopy features, to enhance lesion classification. Patients with histologically confirmed premalignant and malignant oral lesions were included. Endogenous fluorescence images and corresponding Raman spectra from dry serum samples were collected from a total of 66 patients. Features extracted from both modalities were integrated (feature fusion) and evaluated using a Gaussian Naïve Bayes classifier. Due to limited sample size, the model was trained and cross-validated using the complete dataset. Diagnostic performance was assessed via accuracy, precision–recall (PR) analysis, and decision boundary visualization. The multimodal fusion approach yielded an accuracy of 86%, with a precision–recall curve area under the curve (AUC) of 0.94, demonstrating robust discrimination between malignant and premalignant lesions. Decision boundary visualization illustrated effective nonlinear separation, underscoring the complementary nature of combined imaging and biochemical features. Integrating VELscope endogenous fluorescence imaging and dry serum Raman spectroscopy provides enhanced diagnostic accuracy for oral lesion classification. This multimodal feature fusion approach demonstrates potential clinical utility, particularly within small, high-dimensional biomedical datasets, warranting further validation in larger patient cohorts. • A multimodal framework combining VELscope autofluorescence imaging and serum Raman spectroscopy was developed. • Quantitative features were extracted from both modalities and classified using a Naïve Bayes model. • The combined model achieved high precision and recall for both classes (0.88 for malignant, 0.83 for premalignant). • The integrated approach outperformed single-modality models in discriminating lesion types. • The method offers a non-invasive, label-free solution for early detection of oral cancer.

Source: Photodiagnosis and Photodynamic Therapy

DOI: 10.1016/j.pdpdt.2025.105247

Authors: Mukta Sharma, Chia-Lung Tsai, M.V. Arularasu, Shiang-Fu Huang, Yu-Li Hsieh, Liann-Be Chang