| 摘要: |
| 目的：构建一种基于YOLOv4模型的早孕期超声颜面部超声筛查切面自动识别及分数质量评价系统。方法：以妊娠11～14周早孕期颜面超声筛查切面为研究对象，含模型构建集1001张和临床验证集610张。模型构建包括训练集、测试集、验证集，用于训练和验证AI模型对正中矢状面（MSP）、鼻后三角切面(RNT)的分类与识别。以专家级超声医生分类为标准，比较分析AI与专家分类的一致性。结果：模型对测试集图片中各关键解剖结构的查准率、查全率和F1分数均达80%以上；模型对测试集识别性能与专家分类一致性强，MSP、RNT kappa值分别为0.888、0.810；模型对临床验证集图像识别与专家分类一致性良好，MSP、RNT kappa值分别为0.781、0.690；模型对临床验证集图像等级评估与专家评估一致性良好，MSP、RNT Kendall’s W分别为0.760、0.789。结论：AI对早孕期颜面部超声筛查切面分类具有较高准确性，可作为初级及基层医生的培训和图像质量评价的辅助方法。 |
| 关键词: 超声，早孕期，筛查切面，颜面部，人工智能，质量控制 |
| DOI： |
| 投稿时间：2023-01-31修订日期：2023-09-01 |
| 基金项目:福建省自然科学基金资助:2021J011404；泉州市科技计划资助:2022NS057；福建省自然科学基金项目（2023J011773） |
|
| To construct a quality evaluation system for facial ultrasound screening in the first trimester |
| YU Weifeng,XUE hao,LÜ Guorong,LIU Zhonghua,LIU Peizhong,GUO Xu,WU Xiuming |
| (Department of Ultrasound,Quanzhou First Hospital Affiliated to Fujian Medical University) |
| Abstract: |
| Objective To construct an automatic recognition and score quality evaluation system for facial screening in early pregnancy based on the YOLOv4 model. Methods The facial screening images during 11～14 gestational weeks were selected as the research objects, including 1001 for model construction set and 610 for clinical verification set. The model construction included a training set, a test set and a validation set, which were used to train and verify the classification and recognition of the AI model on the midsagittal plane and the retronasal triangle. Taking expert photographer classification as the standard, the consistency between AI and expert classification was compared and analyzed. Results The precision, Recall, and F1-score of the model for the key anatomical structures in the test set were all above 80%. The recognition performance of the model on the test set was extremely consistent with the expert classification, with MSP and RNT kappa values of 0.888 and 0.810, respectively. The model was in favorable agreement with the expert classification in the clinical validation set, with MSP and RNT kappa values of 0.781 and 0.690, respectively. The image grade evaluation of the clinical validation set by the model was in perfect agreement with the expert evaluation, and the MSP and RNT Kendall''s W were 0.760 and 0.789, respectively. Conclusions AI has a strong accuracy in the classification of views for first-trimester screening, and can be used as an auxiliary method for the training and image quality evaluation of primary and primary care doctors. |
| Key words: ultrasound, early pregnancy, screening section, facial, artificial intelligence, Quality control |
