文献速递口腔颌面外科口腔黏膜病

深度神经网络对口腔良恶性病变的分类与诊断

Malignant and non-malignant oral lesions classification and diagnosis with deep neural networks.
PubMed 文献 · J Dent · 2023-08-12

本研究从口腔颌面外科教科书提取342张已确诊口腔病变图像,按非肿瘤性、良性肿瘤、癌前/恶性三类,用MobileNetV3和EfficientNetV2模型分类。EfficientNetV2验证准确率稳定于0.71,MobileNetV3为0.70;两者对癌前/恶性病变测试准确率分别为80%和82%,对非肿瘤性病变均为64%。

要点

  • AI模型对癌前/恶性病变识别准确率较高(80%~82%)。
  • 非肿瘤性病变分类准确率仅64%,需改进模型与分类系统。
  • 智能手机AI软件有望用于偏远地区口腔癌筛查。

核心结论

  • 概念验证研究显示AI软件在区分恶性口腔病变方面具有潜在准确性,可辅助偏远地区筛查。
  • 非恶性病变分类与图像分类结果间存在差异,需进一步优化模型和分类系统。

研究设计与样本

  • 概念验证研究,图像来源于口腔颌面外科教科书,共342张已确诊口腔病变图像。
  • 病变分为三类:Class 1非肿瘤性病变、Class 2良性肿瘤、Class 3癌前/恶性病变。
  • 使用MobileNetV3和EfficientNetV2模型分析,产出准确率曲线、混淆矩阵和ROC曲线。

主要发现

  • EfficientNetV2验证准确率早期陡升,稳定于0.71。
  • EfficientNetV2测试准确率:非肿瘤性病变64%,癌前/恶性病变80%。
  • MobileNetV3验证准确率逐渐上升,稳定于0.70。
  • MobileNetV3测试准确率:非肿瘤性病变64%,癌前/恶性病变82%。

临床怎么用

  • AI软件可作为口腔癌筛查辅助工具,尤其适用于牙医资源有限地区。
  • 当前模型对非肿瘤性病变分类准确性不足,临床使用需谨慎解读结果。

局限性

  • 图像来源于教科书,非真实临床采集,样本量有限(342张)。
  • 非肿瘤性病变分类准确率仅64%,模型和分类系统需进一步改进。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

OBJECTIVES: Given the increasing incidence of oral cancer, it is essential to provide high-risk communities, especially in remote regions, with an affordable, user-friendly tool for visual lesion diagnosis. This proof-of-concept study explored the utility and feasibility of a smartphone application that can photograph and diagnose oral lesions. METHODS: The images of oral lesions with confirmed diagnoses were sourced from oral and maxillofacial textbooks. In total, 342 images were extracted, encompassing lesions from various regions of the oral cavity such as the gingiva, palate, and labial mucosa. The lesions were segregated into three categories: Class 1 represented non-neoplastic lesions, Class 2 included benign neoplasms, and Class 3 contained premalignant/malignant lesions. The images were analysed using MobileNetV3 and EfficientNetV2 models, with the process producing an accuracy curve, confusion matrix, and receiver operating characteristic (ROC) curve. RESULTS: The EfficientNetV2 model showed a steep increase in validation accuracy early in the iterations, plateauing at a score of 0.71. According to the confusion matrix, this model's testing accuracy for diagnosing non-neoplastic and premalignant/malignant lesions was 64% and 80% respectively. Conversely, the MobileNetV3 model exhibited a more gradual increase, reaching a plateau at a validation accuracy of 0.70. The MobileNetV3 model's testing accuracy for diagnosing non-neoplastic and premalignant/malignant lesions, according to the confusion matrix, was 64% and 82% respectively. CONCLUSIONS: Our proof-of-concept study effectively demonstrated the potential accuracy of AI software in distinguishing malignant lesions. This could play a vital role in remote screenings for populations with limited access to dental practitioners. However, the discrepancies between the classification of images and the results of "non-malignant lesions" calls for further refinement of the models and the classification system used. CLINICAL SIGNIFICANCE: The findings of this study indicate that AI software has the potential to aid in the identification or screening of malignant oral lesions. Further improvements are required to enhance accuracy in classifying non-malignant lesions.

来源:J Dent,查看原文。版权归原作者所有。

引用

  1. Liyanage V, Tao M, Park JS, et al. Malignant and non-malignant oral lesions classification and diagnosis with deep neural networks. J Dent. 2023. PMID: 37574105. doi:10.1016/j.jdent.2023.104657

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。