Artificial intelligence-based prenatal ultrasound for diagnosing fetal lip and palate during the second trimester

Fetal facial anomalies are among the most common congenital conditions, with cleft lip and palate being the most prevalent. This study aimed to evaluate the clinical value of convolutional neural networks in automatically identifying standard ultrasonic cross-sectional images of the fetal lip and palate during the second trimester.

From September 2021 to December 2022, prenatal sonographers collected dynamic videos of the lip and palate of 700 fetuses at 20-24 weeks of gestation, including 5 standard cross-sectional images, along with nonstandard cross-sectional images and background images. A YOLOv5-based artificial intelligence (AI) model was used as the object detection network. The sonographers manually marked the images of 500 fetuses (450 in the training set and 50 in the validation set). The AI model, a midcareer sonographer, and a junior sonographer were involved in the classification and identification of standard fetal lip and palate ultrasound images in the test set (200 fetuses), and the results were compared with the standard results obtained by the senior prenatal sonographer. The receiver operating characteristic curve was plotted, and the sensitivity, specificity, and accuracy were calculated for the AI model, midcareer sonographer, and junior sonographer.

For the standard coronal section of the nasal lip, the area under the curve (AUC) of the AI model, the midcareer sonographer, and the junior sonographer was 0.971, 0.935, and 0.880, respectively. For the standard midsagittal section of the face, the AUC of the AI model, the midcareer sonographer, and the junior sonographer was 0.988, 0.939, and 0.904, respectively. For the standard upper alveolar ridge section, the AUC of the AI model, the midcareer sonographer, and the junior sonographer was 0.977, 0.840, and 0.824, respectively. For all standard sections, the AI model demonstrated significantly higher AUC values as compared to both the midcareer and junior sonographers (P<0.05).

The AI model demonstrated higher classification efficacy than did the midcareer and junior sonographers and performed more quickly and efficiently.