| 刘晓棠,黄海珊,马微波,黄稚熙,厉进,陈明朗,梁元豪,江瑶,许凯,陈亚岩.深度学习模型用于自动鉴别常规超声所示子宫内膜息肉与黏膜下肌瘤[J].中国医学影像技术,2025,41(12):1954~1959 |
| 深度学习模型用于自动鉴别常规超声所示子宫内膜息肉与黏膜下肌瘤 |
| Deep learning models for automatically differentiating endometrial polyps and submucosal fibroids in conventional ultrasound |
| 投稿时间：2025-05-29 修订日期：2025-12-14 |
| DOI：10.13929/j.issn.1003-3289.2025.12.004 |
| 中文关键词: 子宫内膜 息肉 平滑肌瘤 超声检查 人工智能 |
| 英文关键词:endometrium polyps leiomyoma ultrasonography artificial intelligence |
| 基金项目:广东省基础与应用基础研究基金(2025A1515011404)。 |
| 作者 | 单位 | E-mail | | 刘晓棠 | 深圳市南山区妇幼保健院超声科, 广东 深圳 518067 | | | 黄海珊 | 中山大学软件工程学院, 广东 珠海 519082 | | | 马微波 | 上海健康医学院护理与健康管理学院, 上海 201318 | | | 黄稚熙 | 南方医科大学附属深圳妇幼保健院新生儿科, 广东 深圳 518028 | | | 厉进 | 深圳市南山区妇幼保健院超声科, 广东 深圳 518067 | | | 陈明朗 | 梧州学院广西机器视觉与智能控制重点实验室, 广西 梧州 543002 | | | 梁元豪 | 东莞市妇幼保健院超声科, 广东 东莞 523000 | | | 江瑶 | 南方医科大学附属深圳妇幼保健院超声科, 广东 深圳 518028 | | | 许凯 | 云南大学软件学院, 云南 昆明 650500 | | | 陈亚岩 | 深圳市龙华区妇幼保健院超声科, 广东 深圳 518000 | 905108774@qq.com |
|
| 摘要点击次数: 88 |
| 全文下载次数: 17 |
| 中文摘要: |
| 目的 探索基于深度学习(DL)模型自动鉴别常规超声所示子宫内膜息肉与黏膜下肌瘤的价值。方法 回顾性收集接受经阴道超声检查的1 291人,按9∶1比例随机划分训练集(n=1 166,含572例子宫内膜息、215例子宫黏膜下肌瘤及379名正常对照)与验证集(n=125,含59例子宫内膜息、23例子宫黏膜下肌瘤及43名正常对照)。基于常规超声图像分别以DenseNet-121、DenseNet-201、EfficientNet-B0、VGG-16、MobileNetV2、ResNet-18及ResNet-34和Inception-v3共8种卷积神经网络构建DL模型以自动鉴别子宫内膜息肉、子宫黏膜下肌瘤与正常子宫;以验证集评估各DL模型分类效能。结果 8种模型鉴别子宫内膜息肉、子宫黏膜下肌瘤及正常子宫的平均精确率、召回率、F1分数均为73.00%~86.00%;其中,Inception-v3模型的平均精确率(85.22%)、召回率(80.69%)及F1分数(82.46%)均较高。基于梯度加权类激活映射分析结果显示,Inception-v3模型提取特征过程关注的区域与临床诊断关注的关键区域高度重叠;t-分布随机邻域嵌入可视化结果显示,Inception-v3模型在三分类任务中表征特征能力均较强,整体分类效能最佳。结论 DL模型可用于鉴别常规超声所示子宫内膜息肉与黏膜下肌瘤,尤以Inception-v3模型效能最佳。 |
| 英文摘要: |
| Objective To explore the value of deep learning (DL) models for automatically differentiating endometrial polyps and submucosal fibroids in conventional ultrasound. Methods A total of 1 291 individuals who underwent transvaginal ultrasound examination were retrospectively enrolled and randomly divided into training set (n=1 166, including 572 cases of endometrial polyps, 215 cases of submucosal uterine fibroids and 379 normal controls) and validation set (n=125, including 59 cases of endometrial polyps, 23 cases of submucosal uterine fibroids and 43 normal controls) at the ratio of 9∶1. Based on conventional ultrasonic images, 8 convolutional neural networks, including DenseNet-121, DenseNet-201, EfficientNet-B0, VGG-16, MobileNetV2, ResNet-18, ResNet-34 and Inception-v3 were used to construct DL models for automatic identification of endometrial polyps, submucosal uterine fibroids and normal uterus, and the classification performance of each DL model in validation set was evaluated. Results The mean precision, recall and F1 scores of 8 models for differentiating endometrial polyps, submucous fibroids and normal uterus were all 73.00%—86.00%, among which the mean precision (85.22%), recall (80.69%) and F1 score (82.46%) of Inception-v3 model were all relative higher. Gradient-weighted class activation mapping analysis showed that the regions of interest identified of Inception-v3 model during feature extraction highly overlapped with the key regions of interest in clinical diagnosis. The visualization results of t-distributed stochastic neighbor embedding indicated that Inception-v3 model had a strong feature representation ability in three-classification task and best overall classification performance. Conclusion DL models could be utilized to differentiate endometrial polyps and submucosal fibroids in conventional ultrasound, among which Inception-v3 model demonstrated the best performance. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
