HiFU-MiPNet: High-Frequency Ultrasound Material Identification and Size Estimation Pipeline

A deep learning-based system for material identification of microspheres using high-frequency ultrasound signals. This project implements both traditional machine learning and deep learning approaches for material classification.

Project Structure

HiFU-MiPNet/ ├── data/ # Dataset directory ├── material_classification_ml/ # Traditional ML implementation │ ├── config.yaml # Configuration file │ ├── train_ml.py # Training script │ ├── feature_extraction.py # Feature extraction utilities │ └── output/ # Training outputs and results ├── material_classification_cnn/ # CNN-based implementation │ ├── config.py # Configuration settings │ ├── model.py # CNN model architecture │ ├── train.py # Training script │ ├── predict.py # Prediction script │ ├── preprocessing.py # Data preprocessing utilities │ ├── peak_extract.py # Peak extraction utilities │ ├── visualize.py # Visualization tools │ └── final_model.h5 # Trained model weights ├── size_estimation_mlp/ # Size estimation implementation │ ├── config.yaml # Configuration settings │ ├── train_size_model.py # Training script │ ├── predict_mat_size.py # Combined material and size prediction │ ├── size_model_arch.py # MLP model architecture │ ├── util.py # Utility functions ├── manual_labeling/ # Manual labeling tools │ ├── manual_label.py # Main labeling script │ ├── size_pred_labeling.py # Size prediction labeling │ ├── load_data.py # Data loading utilities │ ├── plot.py # Visualization tools ├── LICENSE # MIT License └── requirements.txt # Project dependencies

Data

The dataset can be downloaded from Google Drive: Download Dataset

After downloading, place the data files in the data/ directory.

Features

• Multiple classification approaches:
  • Traditional Machine Learning (ML) with feature extraction
  • Deep Learning (CNN) with raw signal processing
• Comprehensive signal processing pipeline
• Peak extraction and analysis
• Material classification
• Size estimation capabilities
• Visualization tools for analysis
• Model evaluation metrics

Requirements

For ML Implementation

• Python 3.8+
• scikit-learn
• NumPy
• Pandas
• SciPy
• Matplotlib
• PyYAML

For CNN Implementation

• Python 3.8+
• TensorFlow 2.x
• NumPy
• Pandas
• SciPy
• Matplotlib
• scikit-learn

Installation

1. Clone the repository: bash git clone https://github.com/naviiidz/HiFU-MiPNet.git cd HiFU-MiPNet

2. Install dependencies: ```bash

   For ML implementation

   pip install -r materialclassificationml/requirements.txt

# For CNN implementation and size estimation pip install -r materialclassificationcnn/requirements.txt ```

Usage

Traditional ML Approach

bash cd material_classification_ml python train_ml.py

CNN-based Approach

```bash cd materialclassificationcnn

Training

python train.py

Prediction

python predict.py ```

MLP Size Estimation

The size estimation component uses a Multi-Layer Perceptron (MLP) to predict the size of microspheres: - Trained on manually labeled size data - Uses bounding box information for prediction - Can be used in combination with material classification

``` cd sizeestimationmlp

Training

python trainsizemodel.py

Prediction: Material + Size using bbox

python predictmatsize.py ```

Manual Labeling

The manual labeling component provides tools for: - Labeling ultrasound signals for material classification - Annotating size information for training - Visualizing and validating labels - Generating labeled datasets

``` cd manual_labeling

Start labeling process

python manuallabel.py --csvoutput OUTPUTNAME.csv --category DATASETCATEGORY --filename FILENAME ```

Data Processing

The system processes ultrasound signals through the following steps: 1. Signal preprocessing and normalization 2. Peak extraction and analysis 3. Feature extraction (for ML approach) 4. Balancing 5. Train/test splitting

Model Architectures

Traditional ML

• Feature extraction from ultrasound signals
• Various ML classifiers (configurable)
• Cross-validation and hyperparameter tuning

CNN Model

• 1D convolutional layers
• Batch normalization
• Max pooling
• Dropout for regularization
• Dense layers for classification

License

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

Contact

Navid Zarrabi (Navid.Zarrabi@torontomu.ca)