| 周晓欣,陈俊谕,卢焕冲,赵志丹,陈绍琦.基于超声图像FPN-SENet-FL深度卷积神经网络模型鉴别乳腺良、恶性肿瘤及预测乳腺癌分子分型[J].中国医学影像技术,2024,40(3):372~377 |
| 基于超声图像FPN-SENet-FL深度卷积神经网络模型鉴别乳腺良、恶性肿瘤及预测乳腺癌分子分型 |
| FPN-SENet-FL deep convolutional neural network model for differentiating benign and malignant breast tumors and predicting molecular subtypes of breast cancers based on ultrasonic images |
| 投稿时间：2023-10-09 修订日期：2024-01-05 |
| DOI：10.13929/j.issn.1003-3289.2024.03.011 |
| 中文关键词: 深度学习 乳腺肿瘤 免疫表型分型 超声检查 |
| 英文关键词:deep learning breast neoplasms immunophenotyping ultrasonography |
| 基金项目:2022年汕头市科技计划医疗卫生类别项目(220516096491790)。 |
|
| 摘要点击次数: 248 |
| 全文下载次数: 161 |
| 中文摘要: |
| 目的 探讨FPN-SENet-FL深度卷积神经网络模型基于超声图像鉴别乳腺良、恶性肿瘤及预测乳腺癌分子分型的可行性。方法 回顾性分析273例乳腺癌480幅及41例乳腺良性肿瘤113幅术前超声图像,以之构建数据集,并以7 ∶ 3比例随机纳入训练集或验证集。对训练集数据进行扩增,以训练二分类及五分类任务模型,评估二分类任务模型鉴别乳腺良、恶性肿瘤,以及五分类任务模型鉴别乳腺良性肿瘤与4种不同分子分型乳腺癌的效能;绘制受试者工作特征曲线,计算曲线下面积(AUC)、混淆矩阵及完成各项任务的准确率、精确率、召回率及F1分数。结果 二分类任务模型的准确率、精确率、召回率及F1分数分别为94.71%、91.32%、91.30%和0.913,其AUC为0.976;五分类任务模型分别为71.78%、72.48%、72.11%及0.721,AUC取值范围为0.860~0.976,其识别良性肿瘤的AUC最高(0.976),其次为识别Luminal B型乳腺癌时(0.944)。结论 FPN-SENet-FL深度卷积神经网络模型可辅助超声鉴别乳腺良、恶性肿瘤,且预测Luminal B型乳腺癌效能较佳。 |
| 英文摘要: |
| Objective To explore the feasibility of FPN-SENet-FL deep convolutional neural network models for differentiating benign and malignant breast tumors and predicting molecular subtypes of breast cancers based on ultrasonic images. Methods Totally 480 preoperative ultrasonic images of 273 breast cancer patients and 113 preoperative ultrasonic images of 41 benign breast tumor patients were retrospectively analyzed. The ultrasonic image dataset was constructed, and the images were randomly divided into the training set or validation set at the ratio of 7 ∶ 3. Data augmentations were applied in the training set, based on which a binary task model and a quinary task model were trained, respectively. The diagnostic performance of the binary task model in differentiating benign breast tumors from malignant ones and the quinary task model in identifying benign breast tumors and different molecular subtypes of breast cancers were evaluated with the receiver operating characteristic curve and the area under the curve (AUC), and the confusion matrix, as well as the accuracy, precision, recall rate and F1-score of the tasks were calculated. Results The accuracy, precision, recall rate and F1-score of binary task model was 94.71%, 91.32%, 91.30%, and 0.913, respectively, with AUC of 0.976, of quinary task model was 71.78%, 72.48%, 72.11% and 0.721, respectively, with AUC of 0.860 to 0.976, and the highest AUC (0.976) was noticed in differentiating benign breast tumors from malignant ones, followed (0.944) in differentiating Luminal B breast cancers from others. Conclusion FPN-SENet-FL deep convolutional neural network model might assist ultrasonic differentiation of benign and malignant breast tumors, with high efficacy for predicting Luminal B breast cancers. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
