USPrediction - Ultrasound Muscle Segmentation

Deep learning pipeline for automatic muscle segmentation in ultrasound videos using U-Net with ResNet34 encoder.

Overview

This project implements a segmentation model to identify and segment different muscle groups in ultrasound imaging. The model is trained on annotated ultrasound videos and can predict pixel-wise segmentation masks in real-time.

Key Features:

• U-Net architecture with ImageNet-pretrained ResNet34 encoder
• Subject-based train/validation split to prevent data leakage
• Multi-class segmentation (up to 8 classes including background)
• Comprehensive data augmentation (speckle noise, gamma correction, random crops)
• Video inference with per-frame loss tracking

Reference: Automatic Muscle Segmentation in Ultrasound Videos

Project Structure

US_Project/
├── code/                    # Source code (see code/README.md)
│   ├── models/unet/         # U-Net model, training, inference
│   └── utils/               # Video/image utilities, augmentations
├── dataset/                 # Raw videos and mask videos
│   ├── raw/                 # Raw ultrasound videos
│   └── masks/               # Annotated mask videos
└── visual_results/          # Generated comparison videos

Visual Results

Side-by-side comparison showing Raw Video | Ground Truth Masks | Model Predictions:

recordings_05_enrollment01_multi_playing01_comparison.compressed.mp4

The comparison video includes:

• Frame counter at the top
• Three panels: raw ultrasound, ground truth segmentation, model prediction
• Real-time visualization of model performance

Authors

• Adir Bruchim (adirbru1@gmail.com)
• Eyal Amdur (eyalamdur@gmail.com)

Technion - Israel Institute of Technology