Anomalyze

Overview

Anomalyze is a deep learning framework for automated nuchal translucency (NT) measurement from fetal ultrasound scans acquired during the 11–14 week gestational window. NT thickness is a clinically validated biomarker for early detection of chromosomal abnormalities (e.g., Trisomy 21), congenital heart defects, and other structural anomalies.

Manual NT measurement is highly operator-dependent and subject to inter-observer variability. EchoWomb standardizes this process through a fully automated pipeline, integrating segmentation, classification, and generative augmentation models to achieve consistent, precise, and real-time NT assessment.

This repository provides the complete workflow: from image preprocessing and model training to evaluation, optimization, and deployment for clinical integration.

✨ Key Features

• Segmentation:

  • U-Net, ResU-Net, and Attention U-Net architectures for pixel-accurate delineation of NT boundaries.

• Classification:

  • EfficientNet and Vision Transformers for NT thickness-based risk stratification.

• Preprocessing Pipeline:

  • CLAHE, Gaussian/wavelet denoising, ROI extraction, and artifact suppression to normalize ultrasound inputs.

• Generative Augmentation:

  • GAN-based synthetic image generation to mitigate dataset scarcity and enhance model robustness.

• Optimization & Evaluation:

  • Hybrid loss functions (Dice + BCE), K-fold cross-validation, AUROC, IoU, and Dice Score.

• Deployment:

  • Model pruning, quantization, and compression for edge-device inference and real-time ultrasound integration.

📊 Methodology

1. Data Collection & Preprocessing

• Dataset: First-trimester fetal ultrasound scans (11–14 weeks).
• Normalization: Intensity standardization across heterogeneous ultrasound devices.
• Enhancement: Adaptive histogram equalization, Gaussian/wavelet denoising.
• ROI Extraction: Bounding-box localization of NT region and patch-based segmentation for fine detail.

2. Model Development

• Segmentation Networks:

  • U-Net baseline for medical image segmentation.
  • ResU-Net with residual skip connections for deeper feature learning.
  • Attention U-Net for contextual focus on NT boundaries.

• Classification Networks:

  • EfficientNet (lightweight, parameter-efficient backbone).
  • Vision Transformers (ViTs) for global contextual reasoning on NT thickness.

• Generative Models:

  • GANs trained on ultrasound distributions to augment rare NT cases.

• Attention Mechanisms:

  • Transformer-based feature refinement for improved classification sensitivity.

3. Training & Evaluation

• Loss Functions:

  • Dice Loss + BCE for segmentation stability.
  • Weighted Cross-Entropy for classification imbalance.

• Validation Protocols:

  • K-fold cross-validation.
  • Metrics: AUROC, IoU, Dice Score, Precision-Recall.

• Efficiency Measures:

  • Model pruning and quantization for faster inference.
  • Batch normalization and mixed-precision training.

4. Deployment Pipeline

• Edge Optimization:

  • Deployment-ready models compressed for bedside ultrasound machines.

• Clinical Validation:

  • Benchmarked against sonographer annotations to assess clinical reliability.

• Deployment Formats:

  • Web-based interface, REST APIs, or firmware-level integration into ultrasound devices.

⚙️ Installation

git clone https://github.com/gupta-nu/EchoWomb.git
cd EchoWomb
pip install -r requirements.txt

🚀 Usage

Preprocess Images

python preprocess.py --input data/raw --output data/processed

Train Segmentation Models

python train.py --model unet --epochs 50

Evaluate Models

python evaluate.py --model unet

🏥 Clinical Relevance

• Why NT? Nuchal translucency measurement in the first trimester is one of the most important non-invasive markers for early prenatal anomaly screening.
• Impact of Automation: EchoWomb reduces human error, accelerates workflows, and enhances reproducibility of NT measurements, enabling broader access to early screening in both advanced and resource-limited clinical settings.

📄 License

MIT License