| 胡海涛,贾芳,王晓荣.超声卷积神经网络模型辅助鉴别颈部淋巴结淋巴瘤与不典型反应性增生[J].中国医学影像技术,2025,41(6):903~907 |
| 超声卷积神经网络模型辅助鉴别颈部淋巴结淋巴瘤与不典型反应性增生 |
| Ultrasound convolutional neural network model for assisting distinguishment of cervical lymph node lymphoma and atypical reactive hyperplasia |
| 投稿时间：2024-11-21 修订日期：2025-04-01 |
| DOI：10.13929/j.issn.1003-3289.2025.06.012 |
| 中文关键词: 头颈部肿瘤 淋巴瘤 增生 超声检查 诊断,鉴别 卷积神经网络 |
| 英文关键词:head and neck neoplasms lymphoma hyperplasia ultrasonography diagnosis, differential convolutional neural network |
| 基金项目:新疆医科大学第一附属医院"青年科研启航"专项基金项目(2023YFY-QKQN-67)。 |
|
| 摘要点击次数: 943 |
| 全文下载次数: 231 |
| 中文摘要: |
| 目的 观察超声卷积神经网络(CNN)模型辅助鉴别颈部淋巴结淋巴瘤与不典型反应性增生的价值。方法 回顾性纳入颈部淋巴结淋巴瘤及不典型反应性增生各335例,以其中520例(淋巴瘤及不典型反应性增生各260例)为开发组、150例为验证组(淋巴瘤及不典型反应性增生各75例)。按7 ∶ 3比例于开发组划分训练集(淋巴瘤及不典型反应性增生各182例)与测试集(含淋巴瘤及不典型反应性增生各78例)。于每例选取1个靶淋巴结,纳入其灰阶超声图及CDFI各1幅用于训练、测试CNN模型及验证模型辅助效能。基于训练集超声图像构建并训练AlexNet、VGG16、ResNet18、DenseNet161、EfficientNet-B0 共5种CNN模型并以测试集进行测试,以分类准确率最高者为最佳模型,观察其辅助鉴别颈部淋巴结淋巴瘤与不典型反应性增生的效能。结果 5种CNN模型中,ResNet18模型在测试集的准确率最高(78.21%),为最佳模型;以之辅助鉴别颈部淋巴结淋巴瘤与不典型反应性增生的敏感度、特异度及准确率均高于医师独立诊断(P均＜0.01)。结论 所构建的ResNet18模型可有效辅助鉴别颈部淋巴结淋巴瘤与不典型反应性增生。 |
| 英文摘要: |
| Objective To observe the value of ultrasound convolutional neural network (CNN) model for assisting distinguishment of cervical lymph node lymphoma and atypical reactive hyperplasia. Methods Totally 335 cases of cervical lymph node lymphoma and 335 cases of atypical reactive hyperplasia were retrospectively enrolled, including 520 cases in development group (260 cases of lymphoma and 260 cases of atypical reactive hyperplasia) and 150 cases in validation group (75 cases of lymphoma and 75 cases of atypical reactive hyperplasia). Patients in development group were divided into training set (182 cases of lymphoma and 182 cases of atypical reactive hyperplasia) and test set (78 cases of lymphoma and 78 cases of atypical reactive hyperplasia) at the ratio of 7 ∶ 3. One target lymph node was selected for each case, and one gray-scale ultrasound image and one CDFI were included for training, testing CNN models and verifying the auxiliary efficacy of the models. Based on ultrasound images in training set, 5 CNN models, including AlexNet, VGG16, ResNet18, DenseNet161 and EfficientNet-B0, were constructed and trained for distinguishing cervical lymph node lymphoma and atypical reactive hyperplasia, and the models were tested in test set to screen out the best one with the highest classification accuracy. The efficacy of the best CNN model for assisting distinguishment of cervical lymph node lymphoma and atypical reactive hyperplasia was observed. Results Among 5 CNN models, the accuracy of ResNet18 model in test set was the highest (78.21%), and ResNet18 model was regarded as the best CNN model, its sensitivity, specificity and accuracy for assisting distinguishment of cervical lymph node lymphoma and atypical reactive hyperplasia were all higher than those of independent diagnosis made by ultrasound physicians (all P＜0.01). Conclusion The constructed ResNet18 model could be used to effectively assist differentiating cervical lymph node lymphoma and atypical reactive hyperplasia. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
