Abstract：
OBJECTIVE: To develop a deep learning (DL)-model using convolutional neural networks (CNN) to automatically identify the fetal head position at transperineal ultrasound in the second stage of labor.
METHODS: Prospective, multicenter study including singleton, term, cephalic pregnancies in the second stage of labor. We assessed the fetal head position using transabdominal ultrasound and subsequently, obtained an image of the fetal head on the axial plane using transperineal ultrasound and labeled it according to the transabdominal ultrasound findings. The ultrasound images were randomly allocated into the three datasets containing a similar proportion of images of each subtype of fetal head position (occiput anterior, posterior, right and left transverse): the training dataset included 70 %, the validation dataset 15 %, and the testing dataset 15 % of the acquired images. The pre-trained ResNet18 model was employed as a foundational framework for feature extraction and classification. CNN1 was trained to differentiate between occiput anterior (OA) and non-OA positions, CNN2 classified fetal head malpositions into occiput posterior (OP) or occiput transverse (OT) position, and CNN3 classified the remaining images as right or left OT. The DL-model was constructed using three convolutional neural networks (CNN) working simultaneously for the classification of fetal head positions. The performance of the algorithm was evaluated in terms of accuracy, sensitivity, specificity, F1-score and Cohen\'s kappa.
RESULTS: Between February 2018 and May 2023, 2154 transperineal images were included from eligible participants across 16 collaborating centers. The overall performance of the model for the classification of the fetal head position in the axial plane at transperineal ultrasound was excellent, with an of 94.5 % (95 % CI 92.0--97.0), a sensitivity of 95.6 % (95 % CI 96.8-100.0), a specificity of 91.2 % (95 % CI 87.3-95.1), a F1-score of 0.92 and a Cohen\'s kappa of 0.90. The best performance was achieved by the CNN1 - OA position vs fetal head malpositions - with an accuracy of 98.3 % (95 % CI 96.9-99.7), followed by CNN2 - OP vs OT positions - with an accuracy of 93.9 % (95 % CI 89.6-98.2), and finally, CNN3 - right vs left OT position - with an accuracy of 91.3 % (95 % CI 83.5-99.1).
CONCLUSIONS: We have developed a DL-model capable of assessing fetal head position using transperineal ultrasound during the second stage of labor with an excellent overall accuracy. Future studies should validate our DL model using larger datasets and real-time patients before introducing it into routine clinical practice.
METHODS: Prospective, multicenter study including singleton, term, cephalic pregnancies in the second stage of labor. We assessed the fetal head position using transabdominal ultrasound and subsequently, obtained an image of the fetal head on the axial plane using transperineal ultrasound and labeled it according to the transabdominal ultrasound findings. The ultrasound images were randomly allocated into the three datasets containing a similar proportion of images of each subtype of fetal head position (occiput anterior, posterior, right and left transverse): the training dataset included 70 %, the validation dataset 15 %, and the testing dataset 15 % of the acquired images. The pre-trained ResNet18 model was employed as a foundational framework for feature extraction and classification. CNN1 was trained to differentiate between occiput anterior (OA) and non-OA positions, CNN2 classified fetal head malpositions into occiput posterior (OP) or occiput transverse (OT) position, and CNN3 classified the remaining images as right or left OT. The DL-model was constructed using three convolutional neural networks (CNN) working simultaneously for the classification of fetal head positions. The performance of the algorithm was evaluated in terms of accuracy, sensitivity, specificity, F1-score and Cohen\'s kappa.
RESULTS: Between February 2018 and May 2023, 2154 transperineal images were included from eligible participants across 16 collaborating centers. The overall performance of the model for the classification of the fetal head position in the axial plane at transperineal ultrasound was excellent, with an of 94.5 % (95 % CI 92.0--97.0), a sensitivity of 95.6 % (95 % CI 96.8-100.0), a specificity of 91.2 % (95 % CI 87.3-95.1), a F1-score of 0.92 and a Cohen\'s kappa of 0.90. The best performance was achieved by the CNN1 - OA position vs fetal head malpositions - with an accuracy of 98.3 % (95 % CI 96.9-99.7), followed by CNN2 - OP vs OT positions - with an accuracy of 93.9 % (95 % CI 89.6-98.2), and finally, CNN3 - right vs left OT position - with an accuracy of 91.3 % (95 % CI 83.5-99.1).
CONCLUSIONS: We have developed a DL-model capable of assessing fetal head position using transperineal ultrasound during the second stage of labor with an excellent overall accuracy. Future studies should validate our DL model using larger datasets and real-time patients before introducing it into routine clinical practice.
摘要：
目的：使用卷积神经网络（CNN）开发深度学习（DL）模型，以在第二产程中自动识别经会阴超声检查时的胎儿头部位置。
方法：前瞻性，多中心研究，包括单例，term,在第二产程中的头胎妊娠。我们使用经腹超声评估胎儿头部位置，随后，使用经会阴超声在轴向平面上获得胎儿头部的图像，并根据经腹超声检查结果进行标记。将超声图像随机分配到三个数据集中，这些数据集包含相似比例的胎儿头位置的每种亚型图像（前枕骨，后部,右横向和左横向)：训练数据集包括70%，验证数据集15%，和测试数据集15%的采集图像。预训练的ResNet18模型被用作特征提取和分类的基础框架。CNN1被训练来区分枕前（OA）和非OA位置，CNN2将胎头错位分类为枕骨后(OP)或枕骨横向(OT)位置，CNN3将其余图像分类为右或左OT。DL模型是使用三个同时工作的卷积神经网络（CNN）构建的，用于胎儿头部位置的分类。在准确性方面评估了算法的性能，灵敏度,特异性，F1分数和科恩的卡帕。
结果：在2018年2月至2023年5月之间，纳入了来自16个合作中心的合格参与者的2154张经会阴图像。经会阴超声在轴向平面中对胎儿头部位置进行分类的模型的整体性能非常出色，占94.5%（95%CI92.0--97.0），灵敏度为95.6%（95%CI96.8-100.0），特异性为91.2%（95%CI87.3-95.1），F1评分为0.92，科恩的卡帕为0.90。CNN1-OA位置与胎儿头部错位-准确率为98.3％（95％CI96.9-99.7），其次是CNN2-OP与OT位置-准确率为93.9%(95%CI89.6-98.2)，最后，CNN3-右侧与左侧OT位置-准确率为91.3%(95%CI83.5-99.1)。
结论：我们开发了一种DL模型，能够在第二产程中使用经会阴超声评估胎儿头部位置，具有出色的总体准确性。未来的研究应该在将其引入常规临床实践之前，使用更大的数据集和实时患者来验证我们的DL模型。
方法：前瞻性，多中心研究，包括单例，term,在第二产程中的头胎妊娠。我们使用经腹超声评估胎儿头部位置，随后，使用经会阴超声在轴向平面上获得胎儿头部的图像，并根据经腹超声检查结果进行标记。将超声图像随机分配到三个数据集中，这些数据集包含相似比例的胎儿头位置的每种亚型图像（前枕骨，后部,右横向和左横向)：训练数据集包括70%，验证数据集15%，和测试数据集15%的采集图像。预训练的ResNet18模型被用作特征提取和分类的基础框架。CNN1被训练来区分枕前（OA）和非OA位置，CNN2将胎头错位分类为枕骨后(OP)或枕骨横向(OT)位置，CNN3将其余图像分类为右或左OT。DL模型是使用三个同时工作的卷积神经网络（CNN）构建的，用于胎儿头部位置的分类。在准确性方面评估了算法的性能，灵敏度,特异性，F1分数和科恩的卡帕。
结果：在2018年2月至2023年5月之间，纳入了来自16个合作中心的合格参与者的2154张经会阴图像。经会阴超声在轴向平面中对胎儿头部位置进行分类的模型的整体性能非常出色，占94.5%（95%CI92.0--97.0），灵敏度为95.6%（95%CI96.8-100.0），特异性为91.2%（95%CI87.3-95.1），F1评分为0.92，科恩的卡帕为0.90。CNN1-OA位置与胎儿头部错位-准确率为98.3％（95％CI96.9-99.7），其次是CNN2-OP与OT位置-准确率为93.9%(95%CI89.6-98.2)，最后，CNN3-右侧与左侧OT位置-准确率为91.3%(95%CI83.5-99.1)。
结论：我们开发了一种DL模型，能够在第二产程中使用经会阴超声评估胎儿头部位置，具有出色的总体准确性。未来的研究应该在将其引入常规临床实践之前，使用更大的数据集和实时患者来验证我们的DL模型。
