SelectGA: Adaptive Frame Selection for Gestational Age Estimation from Blind Sweep Fetal Ultrasound Videos

[]

SelectGA is a novel AI framework designed to improve gestational age (GA) prediction from blind sweep ultrasound videos in low-resource healthcare settings. Our approach uses adaptive frame selection to identify the most clinically informative frames from ultrasound sweeps, achieving a 27% improvement in prediction accuracy compared to existing methods.

🎯 Key Features

• Adaptive Frame Selection: Intelligently identifies the most informative frames from blind sweep videos
• Anatomically-Guided Filtering: Uses pretrained object detection to focus on frames containing fetal structures
• Diversity-Based Sampling: Employs clustering to select diverse, non-redundant frames
• Resource-Efficient: Designed specifically for low-resource healthcare environments
• Cross-Center Validation: Tested across multiple geographical locations and equipment types

🚀 Performance Highlights

• 9.60 days Mean Absolute Error (MAE) - 27% improvement over baselines
• 63.9% of predictions within 7-day clinical threshold
• R² = 0.906 correlation with ground truth gestational age
• Consistent performance across 2nd and 3rd trimesters
• Validated on multi-center dataset from diverse resource settings

🛠️ Installation

Prerequisites

• Python 3.8+
• CUDA-capable GPU (recommended)
• 16GB+ RAM

Setup Environment

# Clone the repository
git clone https://github.com/tanya-akumu/selectGA.git 
cd SelectGA

# Create conda environment
conda create -n selectga python=3.10
conda activate selectga

# Install dependencies
pip install -r requirements.txt

📊 Dataset

Our framework was validated on a multi-center fetal ultrasound dataset:

• 1,314 blind sweep videos from 162 study scans
• Two geographical centers with different equipment (Philips Lumify, GE Voluson v8)
• 245,048 total frames across diverse gestational ages
• Standardized blind sweep protocol with 6-10 sweeps per patient

Data Structure

data/
├── videos/
│   ├── center1/
│   │   ├── patient_001_sweep_001.mp4
│   │   └── ...
│   └── center2/
│       ├── patient_001_sweep_001.mp4
│       └── ...
├── annotations/
│   ├── train.csv
│   ├── val.csv
│   └── test.csv
└── metadata/
    └── patient_info.csv

Note: Due to privacy regulations, the original dataset cannot be shared publicly. Please use your own blind sweep ultrasound data following the same format.

🎓 Training

Prepare Your Data

1. Organize your data following the structure above
2. Create annotation CSV files with columns: sweep_video_path, gestational_age_days, patient_id
3. Update the configuration file configs/selectga_config.yaml

Train SelectGA Model

# Train with custom parameters
python train.py \
    --data_path /path/to/your/data \
    --batch_size 16 \
    --learning_rate 1e-4 \
    --num_epochs 200 \
    --num_frames 16 \
    --confidence_threshold 0.25

Training Configuration

Key parameters in selectga_config.yaml:

# Data settings
data_path: "./data"
train_split: 0.6
val_split: 0.2
test_split: 0.2

# Model settings
num_frames: 16  # K frames selected per sweep video
confidence_threshold: 0.25  # Object detection threshold
resnet_pretrained: true

# Training settings
batch_size: 16
learning_rate: 1e-4
num_epochs: 200
early_stopping_patience: 5

# Hardware
device: "cuda"
num_workers: 4

Performance Metrics

The evaluation script computes:

• Mean Absolute Error (MAE)
• Root Mean Square Error (RMSE)
• R² correlation coefficient
• Clinical accuracy (% within 7 and 14 days)
• Per-trimester performance
• Cross-center generalization

🏗️ Model Architecture

SelectGA consists of two main stages:

Stage 1: Adaptive Frame Selection

1. Anatomically Guided (AG) Selector

   • Pretrained Faster R-CNN detects fetal structures
   • Filters frames with confidence > threshold (α = 0.25)

2. Diversity Guided (DS) Selector

   • Extracts feature embeddings of filtered anatomical frames using pretrained CNN
   • Applies K-means clustering (K=16) on feature space
   • Selects representative frame closest to each cluster centroid

Stage 2: Gestational Age Prediction

1. ResNet-50 Feature Extractor (ImageNet pretrained)
2. Weighted Average Attention (WAA) Module
3. Regression Head for final GA prediction in days

📊 Results

Quantitative Results

Method	MAE (days) ↓	RMSE (days) ↓	R² ↑	<7d (%) ↑	<14d (%) ↑
ResNet-50	12.90	17.73	0.851	41.7	66.7
EchoNet	11.49	15.86	0.881	47.2	69.4
ViFi-CLIP	12.41	19.38	0.822	52.8	69.4
SelectGA (Ours)	9.60	14.07	0.906	63.9	69.4

Ablation Study

Components	AG	DS	MAE ↓	R² ↑	<7d (%) ↑
ResNet-50 + WAA	✗	✗	12.90	0.851	41.7
+ AG Selector	✓	✗	10.96	0.866	55.5
SelectGA (Full)	✓	✓	9.60	0.906	63.9

Model checkpoints

SelectGA AG/DG blind sweep Selector: https://drive.google.com/file/d/15HkFFgRYTKGqDKzVyOm4bkHZI_Ypb6Cc/view?usp=sharing GA_predictor: https://drive.google.com/file/d/1MT769lm2wxO6VowVOXvAjCZLueURo8sY/view?usp=sharing

🔬 Research Impact

SelectGA addresses critical challenges in global healthcare:

• Accessibility: Enables GA estimation with minimal training data requirements
• Resource Efficiency: Reduces computational demands during training while improving accuracy
• Clinical Relevance: Achieves clinically acceptable accuracy thresholds
• Scalability: Framework designed for deployment in low-resource settings

📄 Citation

If you use SelectGA in your research, please cite our paper:

@InProceedings{AkuTan_Adaptive_MICCAI2025,
        author = { Akumu, Tanya AND Elbatel, Marawan AND Campello, Victor M. AND Osuala, Richard AND Martin-Isla, Carlos AND Valenzuela, Ignacio AND Li, Xiaomeng AND Khanal, Bishesh AND Lekadir, Karim},
        title = { { Adaptive Frame Selection for Gestational Age Estimation from Blind Sweep Fetal Ultrasound Videos } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15973},
        month = {September},
        page = {3 -- 12}
}

📞 Contact

For questions about the code or collaboration opportunities:

• Email: [[email protected]]

📜 License

This project is licensed under the Creative Commons License - see the LICENSE file for details.

⚠️ Important Notes

• Medical Device Regulation: This research code is not intended for clinical use without proper validation and regulatory approval
• Data Privacy: Ensure compliance with local healthcare data regulations (HIPAA, GDPR, etc.)
• Ethical Use: Please use responsibly and consider the broader implications of AI in healthcare

---

Made with ❤️ for improving global maternal and fetal healthcare accessibility