NT Fetal Anomaly Detection

Overview

This project implements deep learning-based NT measurement from fetal ultrasound images to assist in anomaly detection. It employs U-Net for segmentation and EfficientNet for classification, integrating transformers and GANs for dataset augmentation and enhanced robustness.

Features

• NT Segmentation: U-Net, ResU-Net, Attention U-Net for boundary delineation
• NT Classification: EfficientNet, Vision Transformers for thickness prediction
• Preprocessing: CLAHE, Gaussian denoising, ROI extraction, artifact removal
• Generative Models: GANs for synthetic data augmentation
• Optimization: Dice Loss, BCE, K-fold validation, real-time inference tuning
• Deployment: Optimized for edge devices, clinical ultrasound integration

Methodology

1. Data Collection & Preprocessing

• Dataset: Clinical fetal ultrasound images (11–14 weeks gestation)
• Normalization: Intensity standardization across ultrasound devices
• Enhancement: Adaptive histogram equalization, wavelet denoising
• ROI Extraction: Bounding box localization, patch-based segmentation

2. Model Development

• Segmentation: U-Net, ResU-Net, Attention U-Net
• Classification: EfficientNet, Vision Transformers
• Generative Models: GANs for synthetic NT image augmentation
• Attention Mechanisms: Transformer-based feature enhancement

3. Training & Evaluation

• Loss Functions: Dice Loss + BCE (Segmentation), Weighted CE (Classification)
• Validation: K-fold cross-validation, AUROC, Dice Score, IoU
• Efficiency: Quantization, pruning, inference speed optimization

4. Deployment Pipeline

• Edge Optimization: Model compression for real-time inference
• Clinical Validation: Benchmarking with expert annotations
• Deployment Formats: Web-based, API endpoints, ultrasound firmware integration

Installation

git clone https://github.com/gupta-nu/NT-fetal-anomaly-detection.git
cd NT-fetal-anomaly-detection
pip install -r requirements.txt

Usage

Preprocess Images

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

Train Models

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

Evaluate Models

python evaluate.py --model unet

License

MIT License