Affiliation:
1. Shenzhen Maternity and Child Healthcare Hospital
2. Hunan University
Abstract
Abstract
Background
The fetal four-chamber view is essential in prenatal echocardiography for screening and diagnosing fetal heart disease. Accurate cardiac axis (CAx) and cardiothoracic ratio (CTR) measurements are critical for assessing fetal cardiac position and function. This study developed an AI-based model using nnU-NetV2 to segment the four-chamber view automatically and measure the CAx and CTR.
Methods
High-quality fetal four-chamber view images were collected from our hospital. Images meeting the ISUOG criteria were annotated with critical structures to train an AI-based model. Postprocessing enabled automated CAx and CTR measurements with clinical validation by sonographers with 1, 5, and 10 years of experience. The analyses used Dice coefficients, mIoUs, ICCs, and Bland‒Altman plots in R and Python.
Results
A total of 1083 finely labeled images were used to train the model (867 training/validation images, 216 test images). The model demonstrated high segmentation accuracy (Dice coefficient of 87.11 and mIoU of 77.68). Visual inspection verified smooth contour segmentation. The CAx and CTR measurements were highly concordant between the model and sonographers, especially for the sonographers with ten years of experience (CAx ICC 0.83, CTR ICC 0.81). The Bland‒Altman plots showed high agreement between the model and experienced sonographers.
Conclusion
The AI-based model effectively automated the identification and segmentation of critical structures with robust accuracy. It accurately computed CAx and CTR, exhibiting strong concordance with the findings of the senior sonographer. This suggests that the model can assist in diagnosing fetal congenital heart disease through ultrasound while reducing the workload of sonographers.
Publisher
Research Square Platform LLC