机器学习和深度学习方法识别头颈重建皮瓣静脉功能不全
该研究回顾性收集2018至2024年头颈癌患者临床数据和术后皮瓣照片,分为正常组和静脉功能不全组,构建八种机器学习分类器和三种深度学习模型。共分析576例患者的2575张皮瓣图像,ResNet深度学习模型表现最佳,准确率95.23%,AUC 0.940,可作为术后监测辅助工具。
要点
- 深度学习模型ResNet识别皮瓣静脉功能不全准确率95.23%,AUC 0.940。
- 机器学习中随机森林表现最佳,准确率90.25%,AUC 0.759。
- SHAP分析显示Hue_mean和Green_median为关键特征。
核心结论
- 深度学习模型在识别皮瓣静脉功能不全方面具有良好价值,可作为术后监测的辅助工具。
研究设计与样本
- 回顾性收集2018至2024年本中心头颈癌患者的临床数据和术后皮瓣照片,分为正常组和静脉功能不全组。
- 构建八种机器学习分类器和三种深度学习模型(ResNet、GoogleNet、Densenet),使用SHAP和Grad-CAM进行特征分析和验证。
主要发现
- 共分析576例患者的2575张皮瓣图像,其中正常2010张,静脉功能不全565张。
- 机器学习中随机森林表现最佳,准确率90.25%,AUC 0.759。
- SHAP识别Hue_mean和Green_median为关键特征。
- 深度学习中ResNet表现最优,准确率95.23%,灵敏度84.81%,特异度97.27%,AUC 0.940。
临床怎么用
- 深度学习模型可作为头颈重建术后皮瓣静脉功能不全的辅助监测工具,帮助早期发现。
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
BACKGROUND: Venous insufficiency is a major cause of flap failure in head and neck reconstruction. AI provides a reliable, convenient solution for early detection. METHODS: Clinical data and postoperative flap photos of head and neck cancer patients (2018-2024) at our center were retrospectively collected, categorized into normal and venous-insufficient groups. Eight machine learning classifiers and three deep learning models (ResNet, GoogleNet, Densenet) were built. SHAP and Grad-CAM visualization were used for feature analysis and validation. RESULTS: A total of 2575 flap images from 576 patients (2010 normal, 565 venous-insufficient) were analyzed. Random Forest performed best in machine learning (accuracy 90.25%, AUC 0.759), with SHAP identifying Hue_mean and Green_median as key features. ResNet outperformed in deep learning (accuracy 95.23%, sensitivity 84.81%, specificity 97.27%, AUC 0.940). CONCLUSION: The deep learning model shows good value in identifying flap venous insufficiency, serving as an auxiliary tool for postoperative monitoring.
引用
- He Y, Fang J, Hou L, et al. Identifying Venous Insufficiency in Head and Neck Reconstruction Flaps Using Machine Learning and Deep Learning Methods. Head Neck. 2026. PMID: 41839653. doi:10.1002/hed.70235
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。