文献速递头颈神经CT

颞骨HRCT内耳畸形亚型分类的深度学习模型

Deep learning for subtype classification of inner ear malformations on temporal bone HRCT: Development and multicenter validation.
PubMed 文献 · Eur J Radiol · 2026-07-08

本研究基于多中心9182耳颞骨HRCT数据,开发并外部验证了一种弱监督深度学习框架,用于内耳畸形多分类诊断。外部测试集准确率93.9%、宏F1为88.2%、多分类AUC为0.990。读片研究中模型宏F1为89.9%,优于高年资耳鼻喉科医师的75.7%(Holm校正p=0.0046)和低年资医师的62.9%(p<0.001)。

要点

  • 该模型在外部测试集准确率达93.9%,可用于术前风险分层参考。
  • 模型表现优于高、低年资耳鼻喉科医师,适合亚专科经验有限场景。
  • 模型可生成注意力切片排序与Grad-CAM热力图,提供可解释输出。

核心结论

  • 该弱监督框架在颞骨HRCT上实现了准确且具泛化性的内耳畸形亚型诊断,可作为标准化术前评估的决策支持工具。

研究设计与样本

  • 多中心队列研究,共9182耳,其中内耳畸形3161耳、正常6021耳。
  • 采用改进的基于Transformer的多实例学习架构,对自动提取的颞骨感兴趣区进行一阶段分类。
  • 参考标准为颞骨HRCT诊断,机构审查委员会批准并豁免知情同意。
  • 独立读片研究比较模型与5名高年资及5名低年资耳鼻喉科医师。

关键影像表现/征象

  • 模型利用注意力机制输出贡献最高的三个切片并生成Grad-CAM热力图,用于解释分类依据。

诊断标准与关键数值

  • 外部独立测试集:准确率93.9%(95%CI,92.4%-95.3%),宏F1为88.2%(95%CI,84.3%-91.4%),多分类AUC为0.990(95%CI,0.985-0.995)。
  • 内部验证集:准确率96.5%(95%CI,95.6%-97.4%),宏F1为87.5%(95%CI,83.0%-91.0%),多分类AUC为0.994(95%CI,0.989-0.998)。
  • 读片研究中模型准确率94.0%,宏F1为89.9%;高年资医师平均宏F1为75.7%(Holm校正p=0.0046),低年资医师平均宏F1为62.9%(Holm校正p<0.001)。

临床应用与阅片要点

  • 适用于颞骨HRCT内耳畸形亚型分类,辅助术前风险评估与手术规划。
  • 在亚专科经验有限的医疗环境中可作为标准化诊断参考工具使用。
  • 报告解读时可参考模型输出的高贡献切片和热力图区域。

局限性

  • 研究为回顾性多中心设计,且豁免知情同意。
  • 原文未报告模型在不同亚型间误分类的具体模式。

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)。

英文原文(摘要)· 点击展开

PURPOSE: Accurate subtype classification of inner ear malformations (IEMs) on temporal bone high-resolution computed tomography (HRCT) is important for preoperative risk assessment and surgical planning, but depends heavily on subspecialty expertise. This study aimed to develop and externally validate a deep learning model for multiclass IEM diagnosis on temporal bone HRCT.

METHODS: We developed and externally validated a weakly supervised deep learning framework for one-stage classification of automatically extracted temporal bone regions of interest from HRCT in a multicenter cohort of 9,182 ears (3,161 IEMs and 6,021 normal ears). The framework was based on an improved Transformer-based multiple instance learning architecture. Performance was evaluated on internal and independent external datasets. An independent reader study compared the model with five senior and five junior otolaryngologists. Explainability was assessed using attention-based slice ranking and Grad-CAM. The study was approved by an institutional review board, and informed consent was waived.

RESULTS: On the independent external test set, the proposed model demonstrated strong multiclass performance, with an accuracy of 93.9 % (95 % CI, 92.4 %-95.3 %), a macro-F1 score of 88.2 % (95 % CI, 84.3 %-91.4 %), and a multiclass AUC of 0.990 (95 % CI, 0.985-0.995). Performance was consistent on the internal validation set, with an accuracy of 96.5 % (95 % CI, 95.6 %-97.4 %), a macro-F1 score of 87.5 % (95 % CI, 83.0 %-91.0 %), and a multiclass AUC of 0.994 (95 % CI, 0.989-0.998). In the reader study, the model obtained an accuracy of 94.0 % and a macro-F1 score of 89.9 %, outperforming both senior otolaryngologists (mean macro-F1, 75.7 %, Holm-adjusted p = 0.0046) and junior otolaryngologists (mean macro-F1, 62.9 %, Holm-adjusted p < 0.001). The model also provided interpretable outputs by identifying the top three highest-contributing slices and generating Grad-CAM heatmaps.

CONCLUSION: This weakly supervised framework enabled accurate and generalizable subtype-level diagnosis of IEMs on temporal bone HRCT in a large multicenter cohort and may serve as a practical decision-support tool for standardized preoperative assessment, particularly in settings with limited subspecialty expertise.

来源:Eur J Radiol,查看原文。版权归原作者所有。

引用

  1. Li, Zhao, Hao, et al. Deep learning for subtype classification of inner ear malformations on temporal bone HRCT: Development and multicenter validation. Eur J Radiol. 2026. PMID: 42447770. doi:10.1016/j.ejrad.2026.113059

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。