| 摘要: |
| 目的：探讨基于四腔心切面的超声图像传递学习深度模型定量评价胎儿肺发育成熟程度的价值研究。方法：选择2019年1月至2020年12月于本院接受常规妊娠检查的单胎妊娠者，利用超声成像系统采集胎龄从24周至出生前的胎儿心脏四腔心切面图像数据。采用一种两阶段的传递学习方法，建立设计特定U型网络（U-net）结构的深度学习模型。本研究假设胎儿肺发育成熟程度的发展通常与胎龄成正比。第一阶段，该模型对超声图像中胎儿肺区域的识别进行学习；第二阶段，预训练的深度模型经过训练可以从超声图像的胎儿肺区域准确估算胎龄。结果：本研究共收集316位患者的超声图像数据，将前166位患者用于传递学习第一阶段的模型培训，而后的150位患者用于独立测试。建立的深度模型在胎龄估算上的测试结果与真实值相比的不精确度为1.43±2.01周，其与胎龄真实值的相关系数为0.773（95%CI 0.6893-0.8314，P<0.05），经过传递学习后深度模型的估算结果明显优于传统方法的估算结果（P<0.05）。结论：本研究初步验证了超声图像中胎儿肺纹理信息可以反映胎儿肺发育成熟程度发展假说。胎肺成熟度可以用胎龄估计值表示的经过传递学习后的深度模型输出来表示。 |
| 关键词: 超声图像、胎肺成熟度、纹理分析、传递学习、深度模型 |
| DOI： |
| 投稿时间：2021-03-17修订日期：2021-04-26 |
| 基金项目:上海市科委医学重大项目（09DZ1950300） |
|
| Analysis of the diagnostic efficacy of high-frequency ultrasound combined with fasting blood glucose and glycosylated hemoglobin in the diagnosis of median nerve injury in early diabetes |
| Yao Ling |
| () |
| Abstract: |
| Objective: To explore the value of the ultrasound image transfer learning depth model based on the four-chamber heart view to quantitatively evaluate the fetal lung development and maturity.Methods: Select singleton pregnancies who have undergone routine pregnancy checkups in our hospital from January 2019 to December 2020, and use the ultrasound imaging system to collect four-chamber heart image data of the fetal heart from the gestational age of 24 weeks to before birth. A two-stage transfer learning method is adopted to establish a deep learning model that designs a specific U-net structure. This study assumes that the development of fetal lung maturity is usually proportional to gestational age. In the first stage, the model learns the recognition of the fetal lung region in the ultrasound image; in the second stage, the pre-trained depth model is trained to accurately estimate the gestational age from the fetal lung region in the ultrasound image.Results: In this study, a total of 316 patients’ ultrasound image data were collected. The first 166 patients were used for the model training of the first stage of transfer learning, and the next 150 patients were used for independent testing. The inaccuracy of the established depth model on the gestational age estimation test results compared with the true value is 1.43±2.01 weeks, and its correlation coefficient with the true gestational age value is 0.773 (95%CI 0.6893-0.8314, P<0.05) , After transfer learning, the estimation result of the deep model is significantly better than the estimation result of the traditional method (P<0.05). Conclusion: This study preliminarily verified the hypothesis that the fetal lung texture information in ultrasound images can reflect the developmental maturity of fetal lungs. Fetal lung maturity can be represented by the output of the depth model after transfer learning represented by the estimated gestational age. |
| Key words: Ultrasound image fetal lung maturity texture analysis transfer learning deep model |
