| 许才顼,谢龙洋,黄稚熙,文华轩,安丽婷,韦俊名,陈家希,贺杰,何冠南.基于VMamba模型智能识别产前超声标准切面[J].中国医学影像技术,2025,41(12):1944~1949 |
| 基于VMamba模型智能识别产前超声标准切面 |
| Intelligent recognition of prenatal ultrasound standard sections based on VMamba model |
| 投稿时间：2025-05-14 修订日期：2025-07-19 |
| DOI：10.13929/j.issn.1003-3289.2025.12.002 |
| 中文关键词: 胎儿 深度学习 超声检查,产前 |
| 英文关键词:fetus deep learning ultrasonography, prenatal |
| 基金项目:成都医学院联合科研基金(25LHSFY3-01)、广西高校中青年教师科研基础能力提升项目(2024KY0694)。 |
| 作者 | 单位 | E-mail | | 许才顼 | 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
梧州学院电子信息与人工智能学院, 广西 梧州 543002 | | | 谢龙洋 | 云南大学信息学院, 云南 昆明 650091 | | | 黄稚熙 | 南方医科大学深圳妇幼保健院新生儿科, 广东 深圳 518028 | | | 文华轩 | 南方医科大学深圳妇幼保健院超声科, 广东 深圳 518028 | | | 安丽婷 | 南方医科大学深圳妇幼保健院超声科, 广东 深圳 518028 | | | 韦俊名 | 梧州学院电子信息与人工智能学院, 广西 梧州 543002 | | | 陈家希 | 梧州学院电子信息与人工智能学院, 广西 梧州 543002 | | | 贺杰 | 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
梧州学院电子信息与人工智能学院, 广西 梧州 543002 | | | 何冠南 | 四川省妇幼保健院超声科, 四川 成都 646199 | heguannan811003@126.com |
|
| 摘要点击次数: 88 |
| 全文下载次数: 25 |
| 中文摘要: |
| 目的 观察基于VMamba模型智能识别21种产前超声标准切面的价值。方法 回顾性纳入9 099幅符合产前超声筛查指南要求的标准切面,按7 ∶ 1 ∶ 2比例将其分为训练集、测试集与验证集。构建VMamba模型以识别产前超声标准切面,比较其与其他先进深度学习模型,包括ResNet-50、Swin-Transformer、ConvNeXtV2、RepViT及MSPANet的效能,并结合梯度加权类别映射(Grad-CAM)特征热力图、t-SNE分类效果图筛选效能最优模型。结果 VMamba模型自动识别产前超声标准切面的敏感度、特异度、准确率、精确度、受试者工作特征曲线下面积(AUC)及F1分数均最高,分别为94.54%、99.72%、94.42%、94.57%、0.997及0.944。除经侧脑室水平横切面(TV)、左心室流出道切面(LVOT)、尺桡骨冠状切面(UR)及胫腓骨冠状切面(TF)外,其他标准切面所示特征高响应区域均与关键解剖结构高度重合。VMamba模型在特征二维空间中具有较强类内聚集性和类间可分性。结论 基于VMamba模型用于智能识别21种产前超声标准切面表现优异,具备临床转化潜力。 |
| 英文摘要: |
| Objective To observe the value of intelligent recognition of 21 prenatal ultrasound standard sections based on VMamba model. Methods Totally 9 099 standard sections met requirements of prenatal ultrasound screening guidelines were retrospectively enrolled and divided into training set, test set and validation set at a ratio of 7 ∶ 1 ∶ 2. VMamba model was constructed to identify 21 standard sections of prenatal ultrasound, and its recognition performance was compared with other advanced deep learning models including ResNet-50, Swin-Transformer, ConvNeXtV2, RepViT and MSPANet. Then gradient-weighted class activation mapping (Grad-CAM) feature heat map and t-SNE classification effect diagram were combined to screen the best classification model. Results VMamba model had the highest sensitivity, specificity, accuracy, precision, area under the receiver operating characteristic curve (AUC) and F1 score for automatically recognizing standard sections of prenatal ultrasound, which was 94.54%, 99.72%, 94.42%, 94.57%, 0.997 and 0.944, respectively. Except for the transventricular view (TV), left ventricular outflow tract view (LVOT), ulna and radius coronal view (UR) and tibia and fibula coronal view (TF), high-response regions of the features showed in the other standard sections were highly coincident with the key anatomical structures. VMamba model exhibited strong intra-class aggregation and inter-class separability in two-dimensional space of features. Conclusion VMamba model performed outstandingly in intelligent recognition of 21 standard sections of prenatal ultrasound, which was potential for clinical application and transformation. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
