Fetal Cardiac AI Screening
Automated Four-Chamber View Assessment for Enhanced Prenatal Screening
This AI solution automates the extraction and analysis of four-chamber views (4CV) from fetal cardiac ultrasound videos, calculating key biometric parameters like cardiothoracic area ratio (CTAR), cardiac axis, and cardiac position. It aims to provide technical support for examiners, improve obstetric workflow efficiency, and reduce missed abnormalities in prenatal screening for congenital heart disease (CHD). Clinical comparison showed performance comparable to expert obstetricians, particularly benefiting less experienced examiners by standardizing screening accuracy.
Core Functionality Highlights
Fetal cardiac ultrasound screening is crucial for detecting congenital heart disease (CHD), but detection rates are insufficient globally (approx. 60%) due to a shortage of specialists, high variability in diagnostic performance among examiners, and the difficulty of visualizing the small fetal heart. Manual measurements of critical parameters like CTAR, cardiac axis, and point P are highly dependent on examiner experience and skill, leading to potential missed abnormalities and inconsistent screening.
- Automatic 4CV Extraction from ultrasound videos using YOLOv7.
- Segmentation of heart, ventricular septum, whole thorax, and descending aorta using UNet 3+ and SegFormer.
- Automated calculation of cardiothoracic area ratio (CTAR).
- Automated calculation of cardiac axis and fetal position detection.
- Automated assessment of cardiac position (Point P) relative to thoracic centroid.
- Identification of abnormal cardiac orientation and descending aorta position for anomaly detection.
Quantifiable Impact
0.928%
Mean Dice (Heart)
SegFormer mDice for heart segmentation
0.951%
Mean Dice (Whole Thorax)
SegFormer mDice for whole thorax segmentation
2.7%
CTAR MAE
Mean Absolute Error for Cardiothoracic Area Ratio (AI vs. Ground Truth)
5.0°
Cardiac Axis MAE
Mean Absolute Error for Cardiac Axis (AI vs. Ground Truth)
0.860
Expert AUC (Combined)
Area Under Curve for experts combining CTAR, Cardiac Axis, and Point P
0.861
SegFormer AUC (Combined)
Area Under Curve for SegFormer combining CTAR, Cardiac Axis, and Point P
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
0.861
SegFormer AUC (Combined)
SegFormer achieved an Area Under the Curve (AUC) of 0.861 when combining CTAR, cardiac axis, and point P, demonstrating performance comparable to expert obstetricians in fetal cardiac screening.
2.7%
CTAR MAE (AI vs. GT)
Both UNet 3+ and SegFormer achieved a Mean Absolute Error (MAE) of 2.7% for Cardiothoracic Area Ratio (CTAR) compared to ground truth, indicating high accuracy in parameter computation.
0.951%
Whole Thorax mDice
SegFormer achieved a mDice score of 0.951 for whole thorax segmentation, indicating excellent accuracy in delineating the thoracic cavity, crucial for CTAR calculation.
Enterprise Process Flow
Input Fetal Cardiac Ultrasound Video
→
YOLOv7: 4CV Detection & Extraction
→
UNet 3+ / SegFormer: Segmentation (Heart, Thorax, VS, Aorta)
→
Automated Parameter Calculation (CTAR, Cardiac Axis, Point P)
→
Fetal Position & Descending Aorta Confirmation
→
Abnormality Assessment & Screening Outcome
The primary objective of this study was to compare CNN- and Transformer-based segmentation models; therefore, UNet 3+ and SegFormer were selected because they provide high accuracy and inference speed while maintaining low computational cost.
Source: Page 5, Paragraph 1
| Screening Metric | Experts (AUC) | UNet 3+ (AUC) | SegFormer (AUC) |
|---|---|---|---|
| CTAR only | 0.597 | 0.712 | 0.712 |
| CTAR + Cardiac Axis | 0.816 | 0.835 | 0.851 |
| CTAR + Cardiac Axis + Point P | 0.860 | 0.841 | 0.861 |
AI models, particularly SegFormer, demonstrated comparable or superior screening performance (AUC values) to expert obstetricians, especially when combining all biometric parameters. This highlights the potential for AI to standardize and improve screening accuracy.
Impact on Less Experienced Examiners
A key finding from the clinical comparison study was the greater variability in 4CV extraction observed among residents compared to experts. AI-assisted biometric parameter calculation and standardized 4CV extraction directly address this variability.
"AI-assisted biometric parameter calculation and standardized 4CV extraction may particularly benefit less experienced examiners, leading to improved prenatal diagnostic rates."
The fully automated AI models can significantly reduce missed abnormalities and standardize screening accuracy, thereby leading to improved prenatal diagnostic rates, especially when used by less experienced examiners. This reduces the dependency on manual expertise and promotes consistency.
The accuracy of manual measurements of these indices depends on the experience and skill levels of the examiners.
Source: Page 2, Paragraph 3
While AI offers significant advantages, limitations include dependence on ultrasound device quality, generalizability to different gestational ages, and the challenge of low-quality images. Interobserver variability in segmentation labels was not assessed, and the optimal number of images for analysis needs further determination. Future work involves architectural modifications to avoid patent infringement and ablation studies.
Advanced ROI Calculator: Quantify Your AI Advantage
Estimate the potential return on investment for integrating Fetal Cardiac AI Screening into your enterprise workflow.
Estimated Annual Cost Savings
$0
Total Annual Hours Reclaimed
0
Seamless Integration: Your AI Implementation Roadmap
Our structured approach ensures a smooth transition and rapid value realization for your enterprise.
Data Collection & Annotation (2–4 weeks)
Gathering and meticulously annotating a diverse dataset of fetal cardiac ultrasound videos for training and validation, ensuring high-quality ground truth labels.
AI Model Training & Optimization (4–8 weeks)
Training YOLOv7, UNet 3+, and SegFormer models on annotated data, fine-tuning hyperparameters, and performing internal validation to achieve high accuracy and speed.
Clinical Validation & Comparison (6–10 weeks)
Conducting a comprehensive clinical comparison study with obstetricians of varying experience levels to evaluate the AI models' performance against human experts and ensure clinical acceptability.
Integration & Deployment (3–6 weeks)
Integrating the validated AI framework into existing ultrasound systems or a dedicated software platform, followed by rigorous testing in a real-world clinical environment.
Post-Deployment Monitoring & Updates (Ongoing)
Continuous monitoring of AI model performance, gathering feedback from clinicians, and performing iterative updates and retraining to maintain and improve accuracy over time.
Ready to Revolutionize Fetal Cardiac Screening?
By automating 4CV extraction and biometric parameter calculation, the AI solution significantly improves the consistency and accuracy of fetal cardiac ultrasound screening, reducing the dependency on examiner expertise. This leads to earlier and more reliable detection of CHD, ultimately improving prenatal diagnostic rates and patient outcomes. It also streamlines obstetric workflows by providing technical support and flagging cases for specialist review, making comprehensive fetal cardiac evaluations more accessible.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat