Objective: To investigate the application value of near-infrared autofluorescence imaging-based convolution neural network (CNN) for automatic recognition of parathyroid gland. Methods: The data of 83 patients who underwent thyroid papillary cancer surgery in the Department of Head and Neck Surgery, Beijing Tongren Hospital, Capital Medical University from August 2020 to March 2022 were retrospectively analyzed, and a total of 725 autofluorescence images of parathyroid gland were collected during the surgery. Meanwhile, non-parathyroid fluorescence imaging videos in the operation area of 10 patients were also collected, and 928 non-parathyroid fluorescence images were captured from those videos. The fluorescence images of parathyroid and non-parathyroid glands were directly used as input features for deep learning to construct ResNet 34, VGGNet 16 and GoogleNet models for automatic parathyroid identification. The ability of different models to identify parathyroid glands was tested by indicators such as accuracy, specificity, sensitivity, precision, receiver operating characteristic curve and area under the curve (AUC). In addition, 30 fluorescence images of parathyroid and 35 fluorescence images of non-parathyroid glands in 13 patients with papillary thyroid cancer from March to May 2022 were collected to prospectively test the best performing CNN model. Results: Among the 83 patients, there were 25 males and 58 females, with the mean age of (46.7±12.4) years. In the binary classification (parathyroid gland and non-parathyroid gland), the ResNet 34 model performed the best in different CNN models, the accuracy, specificity, sensitivity and precision of the identification test set were 97.6%, 96.3%, 99.3% and 95.5%, and the AUC reached 0.978 (95%CI: 0.956-0.991). In the prospective test, the prediction accuracy of the ResNet 34 model reached 93.8%, and the AUC was 0.938 (95%CI: 0.853-0.984). Conclusion: The near-infrared autofluorescence imaging-based deep CNN has good application value in the automatic recognition of parathyroid gland, and can be used to assist the recognition and protection of parathyroid gland in thyroid cancer surgery.
目的： 探讨基于近红外自体荧光显像的卷积神经网络（CNN）自动识别甲状旁腺的应用价值。 方法： 回顾性分析2020年8月至2022年3月在首都医科大学附属北京同仁医院耳鼻咽喉头颈外科接受甲状腺乳头状癌手术的83例患者临床资料，收集其术中甲状旁腺自体荧光图像共725幅，同时收集其中10例患者术区的非甲状旁腺荧光显像视频，截取非甲状旁腺荧光图像共928幅。将甲状旁腺及非甲状旁腺的荧光图像直接作为深度学习的输入特征，构建用于自动识别甲状旁腺的ResNet 34、VGGNet 16及GoogleNet模型。通过准确率、特异度、灵敏度、精确率、受试者工作特征曲线下面积（AUC）评估不同模型识别甲状旁腺的能力。另外，采集2022年3至5月行甲状腺手术13例患者的甲状旁腺荧光图像30幅，非甲状旁腺荧光图像35幅，对表现最好的CNN模型进行前瞻性测试。 结果： 83例患者中，男25例，女58例，年龄（46.7±12.4）岁。在二分类（甲状旁腺、非甲状旁腺）中，ResNet 34模型在不同的CNN模型中表现最好，识别甲状旁腺的准确率、特异度、灵敏度及精确率分别为97.6%、96.3%、99.3%及95.5%，AUC达到0.978（95%CI：0.956~0.991）。前瞻性测试中ResNet 34模型预测准确率达到93.8%，AUC为0.938（95%CI：0.853~0.984）。 结论： 基于近红外自体荧光显像的深度CNN在自动识别甲状旁腺中具有良好的应用价值，可用于辅助甲状腺癌手术中甲状旁腺的识别和保护。.