甲状腺超声半自动语义标注软件的可重复性评估

Evaluation on repeatability of the semi-automatic semantic annotation software tools for multi-feature thyroid ultrasound images

摘要:

目的：为实现计算机辅助诊断下的甲状腺结节超声图像自动识别，本课题组研发了多特征甲状腺超声图像半自动语文标注软件工具，以完成大样本甲状腺超声图像的高质量人工标注。本研究以探讨超声医生应用甲状腺超声图像标注工具的重复性。方法：随机选取我院2013年甲状腺结节图像50例，分别由2名超声医生独立利用标注软件工具进行甲状腺超声图像识别；2名超声医生独立通过直观超声图像进行超声图像识别；1名超声医生分别利用标注软件工具、直观超声图像进行随机图像识别，并分别利用?资值进行观察者一致性分析。结果：本研究中，2名超声医生分别应用标注软件工具对各项超声特征及超声分级的观察者差异?资值为：边缘/边界0.94，形态0.94，回声0.63，内部结构0.87，钙化0.89，2015版美国甲状腺协会（American Thyroid Association，ATA）指南危险分层0.84（P<0.01）；2名超声医生通过直观超声图像的观察者差异?资值如下：边缘/边界0.75，形态0.80，回声0.53，内部结构0.62，钙化1.00，ATA危险分层0.42（P<0.01）。1名医生分别利用图像标注系统、直观超声图像的观察者差异?资值为：边缘/边界1.00，形态1.00，回声1.00，内部结构0.88，钙化1.00，ATA危险分层0.96（P<0.01）。结论：多特征甲状腺超声图像半自动语义标注软件工具为甲状腺结节超声诊断的标准化提供方法，为其细粒度智能识别研究提供高质量的训练和验证数据集。

Abstract:

Objective：In order to achieve high-quality manual annotation of large-sample thyroid ultrasound images，we developed a semi-automatic semantic annotation software tool for multi-feature thyroid ultrasound images. In the study，we aimed to explore the reproducibility of doctors’ application of the thyroid ultrasound image annotation tool. Methods：Fifty cases of thyroid nodules in our hospital were randomly selected. Two ultrasound doctors independently used the annotation software tool to perform thyroid ultrasound image recognition. Two ultrasound doctors independently performed ultrasound images recognition through visual images.One ultra－sound doctor randomly used annotation software tools and visual images for image recognition，and then observer consistency analysis was performed. Results：In this study，two ultrasound doctors used annotation software tools to observe the differences between ultra－sound features and ultrasound grading（?资 value）：boundary 0.94，morphology 0.94，echo 0.63，internal structure 0.87，calcification 0.89，2015 American Thyroid Association（ATA） risk stratification was 0.84（P<0.01）. The difference between the two ultrasound doctors through the visual ultrasound image was as follows（?资 value）：boundary 0.75，morphology 0.80，echo 0.53，internal structure 0.62，calcification 1.00，ATA risk stratification 0.42（P<0.01）. The difference between the observers using the image annotation system and the visual ultrasound image was（?资 value）：boundary 1.00，morphology 1.00，echo 1.00，internal structure 0.88，calcification 1.00，ATA risk stratification 0.96（P<0.01）. Conclusion：The multi-feature thyroid ultrasound image semi-automatic semantic annota－tion software tool provides a standardized method for the diag－nosis of thyroid nodule ultrasound diagnosis，and provides high-quality training and verification data sets for its fine-grained in－telligent recognition research.