目的：

口腔鳞癌是一种常见且具有较高致死率的恶性肿瘤，早期诊断对于改善患者预后至关重要。传统的病理学方法存在耗时长、依赖医生经验等局限性。近年来，光纤拉曼光谱技术作为一种非侵入性、无标记的快速检测工具，显示出疾病诊断的巨大潜力。然而，关于如何将光纤拉曼光谱技术有效应用于预测口腔鳞癌的肿瘤分期及病理分级，尚需深入研究来验证其可行性和准确性。本研究旨在探讨便携式光纤拉曼光谱技术结合深度学习算法，在口腔鳞癌患者中的应用，即实现口腔鳞癌的分期及病理分级诊断。此外，本研究还利用化学诱导口腔癌小鼠模型，分析在癌变进展中生物分子与光谱的差异，进而解释口腔癌变内在机制。

方法：

本研究设计了一种便携式光纤拉曼光谱仪原型机，用于采集口腔鳞癌患者及小鼠模型的组织光谱数据。招募36例口腔鳞癌患者，按照严格的纳入和排除标准获取样本，并在手术后30分钟内进行光谱数据采集。同时，本研究利用78只C57BL/6小鼠构建不同诱导阶段口腔癌模型，并定期采集舌部组织光谱数据。之后，对所有光谱数据进行预处理，包括去噪、基线校正和归一化。构建基于ResNet50、VGG16及SVM的多任务网络模型，应用用于患者和小鼠的病理预测。使用Grad-CAM方法对模型进行可视化分析，并通过交叉验证评估模型性能。

结果：

在临床研究中，MTN-ResNet50模型在肿瘤分期、淋巴结分期和组织病理分级任务中的准确度分别为94.49 %、94.15 %和94.30 %，特异度和敏感度均接近 95 %。小鼠口腔癌诱导模型研究中，诊断模型在病理分级任务中的准确度、特异度和敏感度分别为91.87 %、95.79 %和91.36 %。拉曼光谱分析表明，在特定的拉曼位移处，不同病理级别和分期的组织之间存在显著的信号差异，这些差异主要涉及糖类、脂质、核酸和胶原蛋白等生物分子。

结论：

便携式光纤拉曼光谱技术结合深度学习算法，能够高效、准确地进行口腔鳞癌分期及病理分级诊断。本研究验证了该技术在临床研究中的适用性，显示出其作为快速、无创诊断工具的巨大潜力。结合深度学习算法及拉曼光谱技术，本研究通过口腔癌诱导小鼠模型解释了口腔癌变过程的内在分子机制。未来工作将致力于设备的集成和优化，开发新型深度学习算法，并扩大临床样本量，以推动该技术在口腔鳞癌诊断中的临床应用。

Objective：

Oral squamous cell carcinoma (OSCC) is a common malignant tumor with a high mortality rate. Early diagnosis is crucial for improving patient prognosis. Traditional pathological methods are limited by being time-consuming and reliant on the physician's expertise. Fiber-optic Raman spectroscopy has emerged as a promising, non-invasive, label-free, and rapid diagnostic tool. However, further research is necessary to verify its feasibility and accuracy in predicting OSCC staging and pathological grading. This study explores the use of portable fiber-optic Raman spectroscopy combined with deep learning algorithms for diagnosing the staging and pathological grading of OSCC. Additionally, it investigates the differences in biomolecules and spectra during cancer progression using a chemically induced oral cancer mouse model to elucidate the underlying mechanisms of oral carcinogenesis.

Methods：

A prototype portable fiber-optic Raman spectrometer was developed to collect tissue spectral data from OSCC patients and mouse models. Thirty-six OSCC patients were recruited following strict inclusion and exclusion criteria, and spectral data were collected within 30 minutes post-surgery. Additionally, 78 C57BL/6 mice were used to create oral cancer models at different induction stages, with tongue tissue spectral data collected periodically. All spectral data underwent preprocessing, including noise reduction, baseline correction, and normalization. A multi-task network model based on MTN-ResNet50, VGG16, and SVM was constructed for pathological prediction in patients and mice. The Grad-CAM method was employed for model visualization, and cross-validation was used to evaluate model performance.

Results：

In clinical studies, the MTN-ResNet50 model achieved accuracies of 94.49%, 94.15%, and 94.30% for tumor staging, lymph node staging, and histopathological grading tasks, respectively, with specificity and sensitivity close to 95%. In the mouse oral cancer induction model study, the diagnostic model achieved accuracies, specificities, and sensitivities of 91.87%, 95.79%, and 91.36%, respectively, for pathological grading tasks. Raman spectroscopy analysis revealed significant signal differences between tissues of different pathological grades and stages at specific Raman shifts, primarily involving biomolecules such as carbohydrates, lipids, nucleic acids, and collagen.

Conclusion:

The combination of portable fiber-optic Raman spectroscopy and deep learning algorithms enables efficient and accurate diagnosis of OSCC staging and pathological grading. This study confirms the applicability of this technology in clinical research, demonstrating its potential as a rapid and non-invasive diagnostic tool. By integrating deep learning algorithms with Raman spectroscopy, this study elucidates the intrinsic molecular mechanisms of oral carcinogenesis using an oral cancer induction mouse model. Future work will focus on device integration and optimization, developing new deep learning algorithms, and expanding clinical sample sizes to promote the clinical application of this technology in OSCC diagnosis.