Not applicable. Not applicable.
Conditions
Interventions
Anomaly detection algorithms designed to automatically distinguish normal from suspicious or inconclusive findings in first-trimester 3D ultrasound data.
Sponsors
Erasmus MC, Universitair Medisch Centrum Rotterdam
Eligibility
Age
16 Years to 99 Years
Inclusion criteria
Inclusion criteria: Inclusion criteria for rater:Gynaecologists /sonographers /prenatal doctors /residents/ clinical researchers with experience in first trimester ultrasound screening or residents /(medical) students working in obstetrics.Working at the Erasmus MC or affiliated through a hospitality agreement.Written informed consent. Inclusion criteria for 3D ultrasound data: Participant in VR FETUS (MEC-2016-541) or VR FETUS 2.0 (MEC-2023-0015)Gestational age between 11+0 and 14+6 weeksAvailable informed consent for reuse of data for future researchAll included datasets were acquired in compliance with ethical standards and with prior approval of the original study. Inclusion criteria for the VR FETUS study were:Women age 18 or above.Viable pregnancy, including multiple pregnancies.Sufficient understanding of the Dutch language (written and spoken).Pregnant women in the first trimester of pregnancy with a high risk of having a fetus with an anomaly (=i.e. the high risk population). Inclusion criteria for VR FETUS 2.0 study are:Women age 16 or above.Viable pregnancy, including multiple pregnancies.Pregnant women in the first trimester of pregnancy (GA = 14+6 weeks) with a high risk of having a fetus with an anomaly (=i.e., the high risk population).Sufficient understanding of the Dutch language (written and spoken).Written informed consent.
Exclusion criteria
Exclusion criteria: Raters exclusion criteria:No written informed consent.Exclusion criteria for 3D ultrasound data: Gestational age <11+0 weeks and >14+6 weeks.No written informed consent concerning the reuse of data for further research.No 3D ultrasounds available or very low quality of 3D ultrasound.VR FETUS or VR FETUS 2. 0 exclusion criteriaVR Fetus exclusion criteria:Women under 18 years of age.Non-viable pregnancy.No sufficient understanding of the Dutch language.Detection of an anomaly in the current pregnancy before randomization. In case of pregnancy duration > 13+6 weeks.VR Fetus 2.0 exclusion criteria: Women under 16 years of age.Non-viable pregnancy.Pregnancy duration > 14+6 weeks.
Design outcomes
Primary
| Measure | Time frame |
|---|---|
| The primary endpoint of this study is the development of deep learning-based AI algorithms for anomaly screening during the first trimester of pregnancy using 3D ultrasound imaging. Algorithm training and validationThe AI algorithms will be trained and validated using annotated data from the VR FETUS and VR FETUS 2.0 databases, incorporating features such as image quality, fetal pose, standard plane identification, biometric measurements, and anatomical landmarks. Performance will be assessed using the following metrics:Accuracy, sensitivity, specificity, precision, recall, and F1-score for classification performance;Offset and angle deviation for quantitative plane localization;Area under the Receiver Operating Characteristic (ROC) curve (AUC) to evaluate overall discrimination ability. Independent testing on unseen dataAfter internal validation, the trained algorithms will be evaluated on an independent subset of previously unseen 3D ultrasound volumes from the VR FETUS database to assess generalizability, reproducibility, and diagnostic reliability. Algorithmic predictions will be compared to manual ground truth annotations using:Sensitivity and specificity, to quantify the algorithm’s ability to correctly identify normal versus suspicious findings;Mean Squared Error (MSE) and Concordance Correlation Coefficient (CCC) for agreement in biometry measurements;Bland–Altman analysis, to evaluate measurement bias and limits of agreement. The training and validation phases will involve splitting the dataset into training, validation, and test subsets. We will perform a cross-validation (k-fold or leave-one-out) to ensure the AI model’s training is consistent, robust and generalizable. The AI model’s performance will be iteratively improved until the results reach saturation and stability in terms of predictive accuracy, at which point external validation with independent data will be conducted. | — |
Secondary
| Measure | Time frame |
|---|---|
| Inter-observer variability A subset of 30 3D ultrasound volumes from the VR FETUS database will be independently annotated by a panel of expert raters. The primary aim is to assess inter-observer variability in plane identification. Statistical analysis of inter-observer variability will be performed using:Intraclass Correlation Coefficient (ICC): To quantify the degree of agreement among the raters for each standard plane.Cohen’s Kappa: To assess the agreement between individual raters for categorical variables (e.g., the presence/absence of anatomical structures). These analyses will provide a benchmark for the AI algorithm's performance by comparing AI-generated annotations to the variability observed between expert raters.We intend to analyze what influence different fetal poses and image quality has on the assessment of the image. Quantitative metrics of plane detectionThe AI algorithms will be evaluated based on quantitative metrics such as offset and angle deviation for the detected planes. These metrics will be compared to those obtained from the expert raters' manual annotations. To assess the clinical relevance of these AI-derived metrics, the following analyses will be performed:Descriptive Statistics: To summarize the distribution of plane offset and angle deviation metrics.Clinically Meaningful Thresholds: Based on the inter-observer variability data, thresholds for optimal and suboptimal plane acquisition will be established. The AI’s performance will then be assessed to determine if the detected planes fall within these thresholds.Correlation Analysis (Pearson/Spearman): To assess the correlation between AI metrics (e.g., plane offset and angle deviation) and clinically relevant parameters, such as biometry measurements and anatomical structure visibility. | — |
Countries
Netherlands
Contacts
Public ContactM. Rousian
Erasmus MC, Universitair Medisch Centrum Rotterdam
Outcome results
None listed