Fig. 3: Dataset selection process.
From: An interpretable deep learning model for first-trimester fetal cardiac screening
Data were collected from seven tertiary centers: Beijing Obstetrics and Gynecology Hospital (BOGH), Maternal and Child Health Hospital of Guangxi Zhuang Autonomous Region (GZAH), Xiangya Hospital of Central South University (XH), Changsha Hospital for Maternal and Child Health Care (CH), Northwest Women’s and Children’s Hospital (NWCH), the Third Affiliated Hospital of Zhengzhou University (TAHZU), and PLA General Hospital (PLAH). The training and validation sets were split at the case level (80 %/20 %) to prevent data leakage from different images of the same case. Cases excluded due to “not readily identifiable on diastolic-phase 4CV” included conditions such as small ventricular septal defects, tricuspid valve dysplasia, or great vessel abnormalities presenting with a normal 4CV appearance. “Incomplete or non-standard 4CV” referred to instances of severe shoulder shadowing, ectopia cordis, or views deemed non-standard by expert review. “Suboptimal quality” exclusions comprised images with insufficient magnification, poor acoustic windows (e.g., due to thick abdominal wall or scars), or excessive Doppler overflow or aliasing. 4CV = four-chamber view.
