利用深度学习算法对胎儿脑超声图像进行正常或异常分类
本研究从大型医院数据库纳入10251例正常和2529例异常妊娠的胎儿脑超声标准轴平面图像,经预处理后获得15372幅正常和14047幅异常图像,按约8:2分为训练集和测试集。深度学习算法用于分割胎儿颅骨、分类图像正常或异常并定位病灶,分割精确率97.9%、召回率90.9%、DICE 94.1%,分类总体准确率96.3%,识别异常图像的敏感度96.9%、特异度95.9%,AUC 0.989(95%CI 0.986-0.991)。算法在61.6%(1535/2491)的异常图像中精确病灶定位,24.6%(614/2491)接近,13.7%(342/2491)不相关。
要点
- 算法识别异常胎儿脑超声图像的敏感度96.9%、特异度95.9%,AUC 0.989(95%CI 0.986-0.991)。
- 病灶定位在61.6%(1535/2491)的异常图像中精确、24.6%(614/2491)接近、13.7%(342/2491)不相关。
- 图像分割精确率97.9%、召回率90.9%、DICE 94.1%。
核心结论
- 深度学习算法可经训练用于胎儿脑超声标准轴平面图像的分割及正常/异常分类,并能提供病灶定位热图;本研究为进一步研究胎儿颅内异常的鉴别诊断奠定基础。
研究设计与样本
- 研究从大型医院数据库纳入10251例正常和2529例异常妊娠的超声图像,异常病例经新生儿超声、随访检查或尸检确认;经预处理后获得15372幅正常和14047幅异常胎儿脑标准轴平面图像,按病例水平约8:2分为训练集和测试集;算法训练用于沿胎儿颅骨进行图像分割、图像正常/异常分类和病灶定位。
关键影像表现/征象
- 图像分割精确率97.9%、召回率90.9%、DICE 94.1%,DICE用于衡量人工标注与机器分割区域的重叠程度。
- 算法在61.6%(1535/2491)的异常图像中精确病灶定位,24.6%(614/2491)接近,13.7%(342/2491)不相关;由专家对算法热图叠加于分割超声图像上的定位满意度进行评判。
诊断标准与关键数值
- 异常图像分类总体准确率96.3%。
- 识别异常图像的敏感度96.9%、特异度95.9%。
- 受试者工作特征曲线下面积AUC 0.989(95%CI 0.986-0.991)。
临床应用与阅片要点
- 算法可对胎儿脑标准轴平面超声图像进行正常/异常分类,并提供热图用于病灶定位,作者认为本研究为进一步研究胎儿颅内异常的鉴别诊断奠定基础。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVES: To evaluate the feasibility of using deep-learning algorithms to classify as normal or abnormal sonographic images of the fetal brain obtained in standard axial planes.
METHODS: We included in the study images retrieved from a large hospital database from 10 251 normal and 2529 abnormal pregnancies. Abnormal cases were confirmed by neonatal ultrasound, follow-up examination or autopsy. After a series of pretraining data processing steps, 15 372 normal and 14 047 abnormal fetal brain images in standard axial planes were obtained. These were divided into training and test datasets (at case level rather than image level), at a ratio of approximately 8:2. The training data were used to train the algorithms for three purposes: performance of image segmentation along the fetal skull, classification of the image as normal or abnormal and localization of the lesion. The accuracy was then tested on the test datasets, with performance of segmentation being assessed using precision, recall and Dice's coefficient (DICE), calculated to measure the extent of overlap between human-labeled and machine-segmented regions. We assessed classification accuracy by calculating the sensitivity and specificity for abnormal images. Additionally, for 2491 abnormal images, we determined how well each lesion had been localized by overlaying heat maps created by an algorithm on the segmented ultrasound images; an expert judged these in terms of how satisfactory was the lesion localization by the algorithm, classifying this as having been done precisely, closely or irrelevantly.
RESULTS: Segmentation precision, recall and DICE were 97.9%, 90.9% and 94.1%, respectively. For classification, the overall accuracy was 96.3%. The sensitivity and specificity for identification of abnormal images were 96.9% and 95.9%, respectively, and the area under the receiver-operating-characteristics curve was 0.989 (95% CI, 0.986-0.991). The algorithms located lesions precisely in 61.6% (1535/2491) of the abnormal images, closely in 24.6% (614/2491) and irrelevantly in 13.7% (342/2491).
CONCLUSIONS: Deep-learning algorithms can be trained for segmentation and classification of normal and abnormal fetal brain ultrasound images in standard axial planes and can provide heat maps for lesion localization. This study lays the foundation for further research on the differential diagnosis of fetal intracranial abnormalities. Copyright © 2020 ISUOG. Published by John Wiley & Sons Ltd.
引用
- Xie HN, Wang N, He M, et al. Using deep-learning algorithms to classify fetal brain ultrasound images as normal or abnormal. Ultrasound Obstet Gynecol. 2020. PMID: 31909548. doi:10.1002/uog.21967
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。