文献速递头颈外科咽喉与嗓音

临床启发的深度学习框架用于喉癌检测与恶变风险评估

A Clinically Inspired Deep Learning Framework for Laryngeal Cancer Detection and Malignancy Transformation Risk Assessment.
PubMed 文献 · Otolaryngol Head Neck Surg · 2025-12-07

本研究基于中国五家医疗中心2585例患者的16587张喉镜图像构建深度学习模型TransCL,用于喉部病变检测、活检引导及良性病变恶变风险预测;内部验证AUC达0.978,外部验证表现稳健,诊断性能与高年资喉科医生相当。

要点

  • TransCL在内部验证中AUC为0.978,准确率0.942,F1分数0.961。
  • 模型对未见病变类型展现强泛化能力,并可预测良性病变恶变风险。
  • 诊断性能优于低年资医生,与顶尖高年资医生相当。

核心结论

  • TransCL在喉癌检测和病变风险评估中表现出高稳健性、准确性和泛化性,诊断性能与经验丰富的临床医生相当,具备临床应用潜力。

研究设计与样本

  • 回顾性分类网络研究,纳入中国五家医疗中心2585例患者的16587张喉镜图像,涵盖10种喉部病变。

主要发现

  • 内部验证AUC为0.978(0.974-0.982),准确率0.942(0.938-0.945),F1分数0.961(0.959-0.963),MCC为0.843(0.831-0.855)。
  • 外部验证性能稳健,优于现有模型。
  • 模型对未见病变类型泛化能力强,能可靠预测良性病变的恶变风险。
  • 与人类专家相比,TransCL优于低年资喉科医生和部分高年资喉科医生,与顶尖高年资医生性能相当。

临床怎么用

  • 可辅助喉镜图像解读,提高病变检测准确性,指导活检决策,优化诊断流程。

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

OBJECTIVE: Laryngeal cancer (LCA) is a prevalent head and neck malignancy. However, laryngoscopic image interpretation remains experience-dependent, resulting in diagnostic variability, particularly in resource-limited settings. This study aimed to develop a deep learning model to enhance lesion detection, guide biopsies, and streamline diagnostic processes. STUDY DESIGN: Retrospective study of a classification network. SETTING: Five medical centers in China. METHODS: A dataset comprising 16,587 laryngoscopic images from 2585 patients representing 10 distinct laryngeal lesions was established. First, TransCL was developed and validated on multiple independent test sets in comparison with existing models. Second, the model was evaluated on unseen lesion types to assess its generalizability. Third, its ability to predict malignant transformation in benign lesions was assessed against expert annotations. Finally, TransCL predictions were compared with diagnoses made by laryngologists. RESULTS: TransCL demonstrated outstanding performance across all test sets, achieving an area under the curve of 0.978 (0.974-0.982), accuracy of 0.942 (0.938-0.945), F1-score of 0.961 (0.959-0.963), and Matthews Correlation Coefficient of 0.843 (0.831-0.855) in the internal validation. Performance remained robust in external validations, outperforming existing models. The model demonstrated strong generalization to unseen lesion types and reliably predicted malignant transformation risk in benign lesions. Compared with human experts, TransCL outperformed junior laryngologists and several senior laryngologists, achieving performance achieving performance comparable to that of top senior laryngologists. CONCLUSIONS: TransCL demonstrated high robustness, accuracy, and generalizability, with diagnostic performance comparable to that of experienced clinicians in LCA detection and lesion risk assessment, suggesting its potential for clinical applications.

来源:Otolaryngol Head Neck Surg,查看原文。版权归原作者所有。

引用

  1. Kang Y, Hu B, Xu K, et al. A Clinically Inspired Deep Learning Framework for Laryngeal Cancer Detection and Malignancy Transformation Risk Assessment. Otolaryngol Head Neck Surg. 2025. PMID: 41353738. doi:10.1002/ohn.70075

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。