FetalVisionAI: A Multimodal AI-Assisted Prenatal Screening System for Fetal Congenital Heart Disease

FetalVisionAI is an AI screening system that integrates fetal ultrasound images and clinical data. It achieves intelligent prenatal detection of congenital heart disease through FetalCLIP embedding, CARDIUM feature fusion, and model calibration technologies. This system aims to address issues such as fetal heart ultrasound examinations relying on physician experience, diagnostic accuracy being affected by subjective factors, and uneven distribution of high-quality medical resources, providing objective and quantifiable auxiliary decision support for clinical practice.

Congenital Heart Disease (CHD) is a common congenital malformation, affecting approximately 8-10 out of every 1000 live births. Early prenatal screening is crucial for intervention and prognosis. However, existing fetal heart ultrasound examinations have high requirements for physician experience, diagnostic accuracy is greatly affected by subjective factors, and high-quality medical resources are unevenly distributed. FetalVisionAI uses deep learning to integrate ultrasound images and clinical data, providing intelligent auxiliary decision support.

Multimodal Data Fusion Design

Integrate ultrasound images (videos/static images) and clinical structured data (maternal age, gestational weeks, medical history, etc.), combining image features with clinical context to enhance screening value.

Key Technical Modules

• FetalCLIP Visual Encoder: A CLIP model adapted for the fetal ultrasound domain, pre-trained on large-scale datasets to capture cardiac structural abnormalities and dynamic changes.
• CARDIUM Clinical Feature Engineering: Convert raw clinical indicators into structured risk features such as demographics, medical history, pregnancy indicators, and ultrasound measurements.
• Multimodal Fusion Strategy: Explore early, middle, late fusion, and attention mechanism fusion to find the optimal solution.
• Model Calibration and Threshold Optimization: Ensure reliable probabilities through calibration techniques such as Platt Scaling and Isotonic Regression; use ROC/PR curve optimization and cost-sensitive learning to address class imbalance.

Clinical Workflow Integration

Design a real-time inference architecture (image acquisition → measurement input → risk assessment → report generation), and enhance interpretability through Grad-CAM heatmaps, attention visualization, etc.

Multicenter Data Collection

Relying on multicenter collaboration, collect diverse data covering different equipment manufacturers, operator levels, and gestational week distributions to ensure the model's generalization ability.

Validation Plan

• Time-split validation: Divide training/validation/test sets by collection time to simulate real scenarios
• External validation: Test with independent hospital data
• Reader Study: Compare diagnostic accuracy between AI-assisted and non-AI-assisted physicians
• Prospective validation: Evaluate system performance in actual clinical environments

Data Scarcity

Address the problem of rare positive cases through data augmentation (video/spatial transformation), semi-supervised learning, transfer learning (adult cardiac ultrasound models), and synthetic data generation.

Image Quality Variation

Preprocess through a quality assessment module to automatically identify low-quality frames and prompt re-acquisition.

Cross-Device Generalization

Use domain adaptation technology to learn device-independent features and improve generalization ability on new devices.

Auxiliary Positioning

The system only outputs risk assessments and does not provide definitive diagnoses. High-risk cases require manual review, and physicians retain the right to veto decisions, emphasizing assistance rather than replacing physician decisions.

Fairness Assessment

Evaluate performance by stratifying race, age, and BMI, monitor algorithmic bias, and ensure consistency across regions with different medical levels.

• Improve screening coverage: Reduce reliance on expert experience and help primary medical institutions conduct high-quality screening
• Standardize diagnostic quality: Reduce diagnostic differences between operators
• Optimize resource allocation: AI handles routine screening, allowing experts to focus on difficult cases
• Improve prognosis: Early detection buys time for intrauterine intervention and postnatal treatment

FetalVisionAI has built a clinically usable AI system for fetal CHD screening through multimodal fusion, domain-specific feature engineering, rigorous calibration, and interpretability design. This project demonstrates the path of combining cutting-edge deep learning technology with medical expertise to solve clinical pain points, providing reference value for the fields of medical AI, computer-aided diagnosis, and multimodal learning.