Artificial intelligence model based on the detection of key anatomical structures for identifying standard planes on ultrasound of the first-trimester fetal central nervous system: a multicenter study.

BACKGROUND

Prenatal ultrasound examination at 11+0 to 13+6 weeks' gestation can detect approximately 33-50% of major fetal structural abnormalities. Despite recent advances in artificial intelligence (AI) for fetal ultrasound, its application in examining the first-trimester fetal central nervous system (FCNS) remains limited. This study aimed to develop, validate, and determine the clinical value of an AI FCNS model (FCNS-Model) based on the detection of key anatomical structures with the purpose of identifying FCNS ultrasound standard planes (FCNS-USPs).

METHODS

This study included 2,901 FCNS-USPs images from three hospitals, which were split into a training set (2,124 images), a test set (264 images), an internal validation set (254 images), and a clinical validation set (259 images). Additionally, an independent external validation set (300 images) was formed from images acquired at a fourth hospital. The performance of junior, intermediate, and senior physicians, as well as that of the FCNS-Model, was compared against the standard of an expert ultrasound team.

RESULTS

In the test set, the model achieved precision, recall, and F1 scores above 92%. For the internal validation set, classification accuracy for each plane was 99.3-100%, with kappa values >0.90, indicating strong agreement with the expert team. In the clinical validation set, FCNS-Model achieved significantly higher areas under the curve (AUCs) than did the senior physician for the anterior-posterior midsagittal plane (0.984 vs. 0.906; P<0.001), the transverse plane through the thalamus (0.982 vs. 0.938; P=0.003), and the coronal plane through the frontal lobe (0.915 vs. 0.852; P=0.03). For the transverse plane of the lateral ventricle, the FCNS-Model's AUC of 0.979 was slightly higher than that of the junior physician (AUC =0.971; P=0.042). Furthermore, McNemar testing demonstrated that the FCNS-Model achieved significantly higher sensitivity than did the junior physician in identifying FCNS-USPs (P=0.02). In the external validation set, the model maintained stable performance despite a moderate decrease, supporting its generalizability.

CONCLUSIONS

FCNS-Model accurately identifies FCNS-USPs and is an effective tool for assisting ultrasound physicians in obtaining standard planes and identifying key anatomical structures, particularly in settings with varying levels of operator experience.