Evaluating the efficacy and diagnostic reliability of computer vision artificial intelligence algorithm in automating anatomical classification for second-trimester fetal anomaly ultrasound scans: A prospective cohort study.
Authors
Affiliations (3)
Affiliations (3)
- Obstetrics and Gynecology Department, Faculty of Medicine, Zagazig University, Zagazig, Egypt.
- Obstetrics and Gynecology Department, Armed Forces Hospitals Southern Region, Khamis Mushait, Saudi Arabia.
- Obstetrics and Gynecology Department, Faculty of Medicine, Alexandria University, Alexandria, Egypt.
Abstract
To evaluate the diagnostic accuracy and workflow efficiency of BioticsAI-anatomyUNet-0.1-2022 software in identifying 18 standard fetal anatomical planes during mid-trimester ultrasound, and to assess the influence of maternal body mass index (BMI) on performance. This prospective cohort study was conducted at Alexandria University Maternity Hospital between May 2023 and April 2025. Singleton pregnancies between 18 and 26 weeks of gestation underwent routine anomaly scans. Images were assessed by BioticsAI-anatomyUNet-0.1-2022, and fetal medicine experts validated its detection results for 18 ISUOG-defined standard planes. Diagnostic metrics (sensitivity, specificity, positive predictive value [PPV], negative predictive value [NPV], and accuracy) were computed per participant and per plane. Linear regression models and one-way analysis of variance (ANOVA) were used to assess the relationship between maternal BMI and software performance. A total of 772 participants contributed 11 823 scan planes. Overall, the software achieved 89.9% sensitivity, 77.4% specificity, 97.5% PPV, 56.9% NPV, and 90.6% accuracy. Femur length and abdominal views achieved the highest accuracy (>98%). In contrast, the cardiac and craniofacial planes demonstrated reduced sensitivity and NPV, particularly for the three vessels and trachea (sensitivity, 62.7%; NPV, 30.8%). The reporting was 44% faster than manual reporting (4.02 vs 7.16 min, P < 0.001). Linear regression showed a statistically significant but clinically modest inverse association between BMI and sensitivity (B = -0.002, R<sup>2</sup> = 0.002, P < 0.001). BioticsAI-anatomyUNet-0.1-2022 demonstrated high diagnostic accuracy and improved workflow efficiency in anomaly scan interpretation. Its performance was robust across BMI levels, though caution is advised for complex planes such as 3VT and craniofacial views.