Rapid multi-task diagnosis of oral cancer leveraging fiber-optic Raman spectroscopy and deep learning algorithms.

Introduction: Oral cancer, a predominant malignancy in developing nations, represents a global health challenge with a five-year survival rate below 50%. Nonetheless, substantial reductions in both its incidence and mortality rates can be achieved through early detection and appropriate treatment. Crucial to these treatment plans and prognosis predictions is the identification of the pathological type of oral cancer. Methods: Toward this end, fiber-optic Raman spectroscopy emerges as an effective tool. This study combines Raman spectroscopy technology with deep learning algorithms to develop a portable intelligent prototype for oral case analysis. We propose, for the first time, a multi-task network (MTN) Raman spectroscopy classification model that utilizes a shared backbone network to simultaneously achieve different clinical staging and histological grading diagnoses. Results: The developed model demonstrated accuracy rates of 94.88%, 94.57%, and 94.34% for tumor staging, lymph node staging, and histological grading, respectively. Its sensitivity, specificity, and accuracy compare closely with the gold standard: routine histopathological examination. Discussion: Thus, this prototype proposed in this study has great potential for rapid, non-invasive, and label-free pathological diagnosis of oral cancer.