Diagnostic Study of Head and Neck Metastatic Tumors From Different Primary Sites Based on Stacking Machine Learning Methods

Metastatic tumors of the head and neck (MTHN) typically indicate advanced disease with a poor prognosis, originating from cells that spread from other body parts. Diagnosis generally relies on slow and error-prone methods like imaging and histopathology. Addressing the need for a faster, more accurate diagnostic method, this study uses hyperspectral imaging to gather detailed cellular data from 208 patients at six primary MTHN sites. Techniques select characteristic spectral bands, and models including SVM, LightGBM, and ResNet are developed. A high-performance classification model, MTHN-SC, employs stacking technology with SVM and LightGBM as base learners and Random Forest as the meta-learner, achieving a diagnostic accuracy of 82.47%, outperforming other models. This research enhances targeted treatment strategies and advances the application of hyperspectral technology in identifying MTHN primary sites.