陶雄杰,邸臻炜,梁博诚,欧阳淑媛,郭慧,贺杰,仝蕊,陈家希,解迪,赵英丽,覃妮,李胜利.基于Swin-Transformer智能辅助模型用于诊断胎儿眼部畸形[J].中国医学影像技术,2025,41(12):1960~1965
基于Swin-Transformer智能辅助模型用于诊断胎儿眼部畸形
Intelligent auxiliary model based on Swin-Transformer for diagnosing fetal ocular malformations
投稿时间:2025-07-26  修订日期:2025-11-28
DOI:10.13929/j.issn.1003-3289.2025.12.005
中文关键词:  畸形  胎儿  超声检查,产前  Swin-Transformer  智能辅助诊断
英文关键词:abnormalities  fetus  ultrasonography, prenatal  Swin-Transformer  intelligent auxiliary diagnosis
基金项目:国家自然科学基金(62162054)、广西自然科学基金(2025GXNSFAA069497)、梧州市科技计划(202302036)、梧州学院校级科研青年项目(2024QN001)。
作者单位E-mail
陶雄杰 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  
邸臻炜 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  
梁博诚 南方医科大学妇女儿童医学中心深圳市妇幼保健院超声科, 广东 深圳 518028  
欧阳淑媛 南方医科大学妇女儿童医学中心深圳市医学遗传中心, 广东 深圳 518028  
郭慧 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  
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  
仝蕊 昆明市妇幼保健院超声医学科, 云南 昆明 650032  
陈家希 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  
解迪 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  
赵英丽 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  
覃妮 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  
李胜利 南方医科大学妇女儿童医学中心深圳市妇幼保健院超声科, 广东 深圳 518028 lishengli63@vip.126.com 
摘要点击次数: 94
全文下载次数: 14
中文摘要:
      目的 观察基于Swin-Transformer的智能辅助模型用于诊断胎儿眼部畸形的价值。方法 回顾性收集经产前筛查确诊眼部畸形胎儿的1 282幅及526幅正常胎儿眼部声像图,按8∶1∶1比例划分训练集、验证集及测试集。基于Swin-Transformer构建智能辅助诊断模型,并与4种主流模型MobileNet-V2、ResNet-50、VGG-16及Vision-Transformer比较其效能。结果 基于Swin-Transformer智能辅助模型诊断测试集胎儿眼部畸形的敏感度为88.31%、特异度为97.37%,受试者工作特征(ROC)曲线的曲线下面积为0.990、精确率为87.31%、F1分数为87.71%,均优于4种主流模型。Swin-Transformer模型在诊断所有畸形的热力图中均呈高度聚焦,混淆矩阵分析显示聚集明显,ROC曲线显示其同时诊断各畸形的效能最佳,t-SNE特征分布聚类边界更清晰且性能稳定。结论 基于Swin-Transformer智能辅助模型用于产前诊断胎儿眼部畸形具有较高准确性与稳定性,有望为辅助诊断胎儿眼部畸形提供关键技术支撑。
英文摘要:
      Objective To observe the value of intelligent auxiliary model based on Swin-Transformer for diagnosing fetal eye malformations. Methods A total of 1 282 ophthalmic ultrasound images of fetuses with ocular malformations confirmed by prenatal screening and 526 ophthalmic ultrasound images of normal fetuses were retrospectively collected and divided into training set, validation set and test set in a ratio of 8∶1∶1. An intelligent auxiliary diagnosis model was built based on Swin-Transformer, and its efficiency for diagnosing fetal eye malformations was compared with that of MobileNet-V2, ResNet-50, VGG-16 and Vision-Transformer models. Results Intelligent auxiliary diagnosis model based on Swin-Transformer achieved a sensitivity of 88.31%, specificity of 97.37%, area under the receiver operating characteristic (ROC) curve of 0.990, precision of 87.31% and F1 score of 87.71% for diagnosing fetal ocular malformations in test set, all superior to those of the other 4 mainstream models. Swin-Transformer model exhibited a high degree of focus in all anomaly diagnosis heat maps, with clear clustering shown by confusion matrix analysis. ROC curve showed that Swin-Transformer model had the best performance in simultaneously diagnosing each anomaly, its t-SNE feature distribution clustering boundary was clear, and the performance was stable. Conclusion Intelligent auxiliary model based on Swin-Transformer exhibited high accuracy and stability for diagnosing fetal eye malformations, promising to provide crucial technical support for the auxiliary diagnosis of fetal ocular malformations.
查看全文  查看/发表评论  下载PDF阅读器