| 张志伟,武杰,边云,田慧,邵成伟.基于迁移学习InceptionV3网络模型鉴别胰腺浆液性囊腺瘤与黏液性囊腺瘤[J].中国医学影像技术,2023,39(6):876~879 |
| 基于迁移学习InceptionV3网络模型鉴别胰腺浆液性囊腺瘤与黏液性囊腺瘤 |
| Transfer learning-based InceptionV3 network model for differentiating pancreatic serous cystic neoplasm and mucinous cystic neoplasm |
| 投稿时间：2023-01-03 修订日期：2023-05-07 |
| DOI：10.13929/j.issn.1003-3289.2023.06.018 |
| 中文关键词: 胰腺肿瘤 深度学习 磁共振成像 |
| 英文关键词:pancreatic neoplasms deep learning magnetic resonance imaging |
| 基金项目: |
|
| 摘要点击次数: 2945 |
| 全文下载次数: 545 |
| 中文摘要: |
| 目的 基于迁移学习建立InceptionV3网络模型，观察其鉴别胰腺浆液性囊腺瘤（SCN）与黏液性囊腺瘤（MCN）的价值。方法 回顾性分析172例胰腺囊性肿瘤（PCN）患者，包括107例SCN和65例MCN；按3 ∶ 1 ∶ 1将其分为训练集（n=102）、验证集（n=34）和测试集（n=36），提取肿瘤T2WI特征。采用迁移学习方法建立InceptionV3网络模型并加以微调，与其他3种卷积神经网络模型和1种机器学习模型进行对比，评估InceptionV3网络模型鉴别胰腺SCN与MCN的价值。结果 基于迁移学习建立的InceptionV3网络模型鉴别胰腺SCN与MCN的准确率为96.13%，敏感度为96.55%，特异度为95.86%，F1值为0.96；所有效能指标均好于非迁移模型。替换全连接层（FC）基于迁移学习InceptionV3模型鉴别胰腺SCN与MCN的准确率为97.13%，敏感度为95.32%，特异度为98.79%，F1值为0.97，均优于冻结卷积层（Conv）及替换FC+冻结Conv模型。基于迁移学习InceptionV3网络模型鉴别胰腺SCN与MCN的效能指标均优于其他4种模型。结论 基于迁移学习成功建立的InceptionV3网络模型对于鉴别胰腺SCN与MCN具有一定价值。 |
| 英文摘要: |
| Objective To establish a transfer learning-based InceptionV3 network model, and to observe its value for differentiating pancreatic serous cystic neoplasm (SCN) and mucinous cystic neoplasm (MCN). Methods Data of 172 patients with pancreatic cystic neoplasm (PCN) including 107 cases of SCN and 65 cases of MCN were retrospectively analyzed. The patients were divided into training set (n=102), verification set (n=34) or test set (n=36) at the ratio of 3:1:1. Then tumor T2WI features were extracted, and an InceptionV3 network model was established with transfer learning method and finely tuned, its value for differentiating pancreatic SCN and MCN was analyzed through comparison with other 3 convolutional neural network models and 1 machine learning model. Results The accuracy of transfer learning-based InceptionV3 network model for differentiating pancreatic SCN and MCN was 96.13%, with sensitivity of 96.55% and specificity of 95.86%, and the F1 value was 0.96, all were superior to those of non-transfer learning model. The accuracy, sensitivity, specificity and F1 value of transfer learning-based InceptionV3 network model that replaced the full connection layer (FC) for differentiating pancreatic SCN and MCN was 97.13%, 95.32%, 98.79% and 0.97, respectively, all were superior to those of frozen convolutional layer (Conv) and replaced FC+frozen Conv models. The performance indicators of transfer learning-based InceptionV3 network model for differentiating pancreatic SCN and MCN was superior to other 3 convolutional neural network models and 1 machine learning model. Conclusion The successfully established transfer learning-based InceptionV3 network model had a certain value for differentiating pancreatic SCN and MCN. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
