【摘要】 目的:基于深度学习算法，对ViT-B模型检测口腔良性和恶性病变图像的性能进行分析，旨在为临床医生早期发现和准确诊断口腔癌提供有效工具。方法:使用包含口腔良性和恶性病变图像的公共数据集，对数据进行预处理和数据增强，按7∶2∶1的比例将数据随机划分为训练集、验证集和测试集。选取ViT-B、VGG16、ResNet101、DenseNet121和EfficientNetV2 5种深度学习模型，对模型进行训练和性能比较。通过外部数据对ViT-B模型的泛化能力进行评估，并基于注意力权重的可视化方法对ViT-B模型进行分析。结果:ViT-B在5种模型中分类性能最佳，受试者工作特征曲线下面积为0.9715,准确率为91.00%。该模型可以有效区分口腔良性和恶性病变图像，具有较强的泛化能力和临床实用性。结论:ViT-B模型在口腔良性和恶性病变图像识别中表现良好，可以为口腔癌的早期发现和准确诊断提供支持。更多还原

【Abstract】 Objective: To analyze the performance of ViT-B model in detecting oral benign and malignant lesions based on deep learning algorithms. Methods: A public dataset containing images of oral benign and malignant lesions was used, with preprocessing and data augmentation applied. The data was randomly divided into training, validation, and test sets in a 7∶2∶1 ratio. Five deep learning models, including ViT-B, VGG16, ResNet101, DenseNet121, and EfficientNetV2, were selected for training and evaluation. The generalization ability of the ViT-B model was evaluated using external data, and the model was analyzed based on the visualization of attention weights. Results: The ViT-B model demonstrated the best performance among five models, with an area under the receiver operating characteristic curve(AUC) of 0.9715 and an accuracy of 91.00%. The model effectively distinguished between images of oral benign and malignant lesions, demonstrating strong generalization ability and clinical applicability. Conclusion: The ViT-B model performs well in the recognition of oral benign and malignant lesions, supporting the early detection and accurate diagnosis of oral cancer.更多还原