CNN-PD-3D

Convolutional Neural Network for Automated Parkinson's Disease Detection from Structural 3D MRI

A MATLAB-based deep learning toolbox that uses 3D ResNet-18 transfer learning for binary classification of Parkinson's disease (PD) vs. healthy controls from structural MRI scans.

---

📋 Table of Contents

• Overview
• Features
• Requirements
• Installation
• Usage
  • GUI Application
  • Training Pipeline
  • Inference
• Data Format
• Model Architecture
• File Descriptions
• Citation
• License
• Contact

---

🎯 Overview

CNN-PD-3D is a deep learning framework designed for automated detection of Parkinson's disease from structural 3D MRI brain scans. The tool leverages transfer learning from a pre-trained 3D ResNet-18 architecture, adapted for volumetric neuroimaging data.

Key Contributions

• 3D Transfer Learning: Adapts 2D ResNet-18 weights for 3D volumetric MRI analysis
• Data Augmentation: Implements 3D spatial transformations (rotation, scaling, translation) for robust training
• User-Friendly GUI: Provides an intuitive graphical interface for clinical use
• Two-Stage Training: Supports iterative training to improve model accuracy

---

✨ Features

• 🧠 3D Convolutional Neural Network based on ResNet-18 architecture
• 📊 Transfer Learning from pre-trained ImageNet weights adapted for 3D
• 🔄 Data Augmentation Pipeline with randomized 3D affine transformations
• 🖥️ MATLAB GUI for easy clinical deployment
• 📈 GPU Acceleration for faster training and inference
• 📁 NIfTI Support for standard neuroimaging file formats

---

💻 Requirements

Software

• MATLAB R2020b or later
• Deep Learning Toolbox
• Image Processing Toolbox
• Computer Vision Toolbox (optional, for some augmentations)

Hardware

• RAM: Minimum 16 GB (32 GB recommended)
• GPU: NVIDIA GPU with CUDA support (recommended for training)
• Storage: ~5 GB for model weights and sample data

MATLAB Toolboxes

% Check required toolboxes
ver('nnet')           % Deep Learning Toolbox
ver('images')         % Image Processing Toolbox

---

📥 Installation

1. Clone the repository

   git clone https://github.com/yourusername/CNN-PD-3D.git
   cd CNN-PD-3D

2. Download pre-trained weights

   • Ensure params.mat (pre-trained 3D ResNet-18 weights) is in the root directory
   • The trained PD model 3DPretrainedModel.mat will be generated after training

3. Add to MATLAB path

   addpath(genpath('CNN-PD-3D'));

---

🚀 Usage

GUI Application

Launch the graphical user interface for single-subject inference:

CNN_3D_PD

1. Click the Load Image button
2. Select a NIfTI (.nii) file containing the 3D MRI scan
3. View the probability of Parkinson's disease

Training Pipeline

First-Time Training

For initial model training with your dataset:

% Run the first training pipeline
run('Pipeline3D_for1stTimeONLY_Training.m')

This script will:

• Load MRI data from Dataset_Updated/Processed images - 1st session/
• Initialize 3D ResNet-18 with transfer learning
• Train with data augmentation
• Save the model as 3DPretrainedModel.mat

Iterative Training (Fine-tuning)

To improve accuracy with additional data sessions:

% Run 2-3 times for improved accuracy
run('Pipeline3D_forREST_Training.m')

Training Options

Parameter	Default	Description
initLearnRate	0.001	Initial learning rate
maxEpochs	15	Maximum training epochs
miniBatchSize	Auto	Calculated based on dataset size
valFrequency	4	Validation frequency per epoch

Inference

For programmatic inference on new data:

% Load trained model
load('3DPretrainedModel.mat');

% Set up datastore for test image
inputSize = mriNet.Layers(1).InputSize;
imds = imageDatastore('path/to/image.nii', ...
    'FileExtensions', '.nii', ...
    'ReadFcn', @niftiread, ...
    'ReadSize', 10);

% Preprocess and classify
augimdsTest = transform(imds, @(data)classification3DAugmentationPipeline1(data, inputSize, 'test'));
[label, probs] = classify(mriNet, augimdsTest);

% Display result
fprintf('PD Probability: %.2f%%\n', 100*probs(2));

---

📁 Data Format

Expected Directory Structure

Dataset_Updated/
├── Processed images - 1st session/
│   ├── Subject001_Healthy.nii
│   ├── Subject002_Healthy.nii
│   ├── ...
│   ├── Subject026_Parkinson.nii
│   └── Subject027_Parkinson.nii
└── Processed images - 2nd session/
    └── [Same structure]

Input Specifications

Specification	Value
File Format	NIfTI (.nii)
Input Size	224 × 224 × 224 × 1
Preprocessing	Z-score normalization
Cropping Region	[7, 6, 6] to [94, 78, 78]

---

🏗️ Model Architecture

3D ResNet-18 Structure

Input Layer (224×224×224×1)
    │
    ▼
Conv3D (7×7×7, 64) → BN → ReLU → MaxPool3D
    │
    ▼
[ResBlock × 2] (64 filters)
    │
    ▼
[ResBlock × 2] (128 filters)
    │
    ▼
[ResBlock × 2] (256 filters)
    │
    ▼
[ResBlock × 2] (512 filters)
    │
    ▼
Global Average Pooling 3D
    │
    ▼
Fully Connected (2 classes)
    │
    ▼
Softmax → Classification Output

Data Augmentation (Training)

• Rotation: ±15°
• Scaling: 0.85× to 1.15×
• Translation: ±15 voxels (X, Y, Z)

---

📄 File Descriptions

File	Description
CNN_3D_PD.m	Main GUI application for PD classification
CNN_3D_PD.fig	MATLAB GUI layout file
Pipeline3D_for1stTimeONLY_Training.m	Initial training pipeline script
Pipeline3D_forREST_Training.m	Iterative fine-tuning pipeline
classification3DAugmentationPipeline.m	Data augmentation for train/val/test
classification3DAugmentationPipeline1.m	Simplified preprocessing for inference
resnet18TL3Dfunction.m	3D ResNet-18 architecture definition
params.mat	Pre-trained 3D ResNet-18 weights
3DPretrainedModel.mat	Trained PD classification model

---

📚 Citation

If you use this tool in your research, please cite:

@software{cnn_pd_3d,
  author = {Negida, Ahmed},
  title = {CNN-PD-3D: Convolutional Neural Network for Automated Parkinson's Disease Detection from Structural 3D MRI},
  year = {2026},
  url = {https://github.com/yourusername/CNN-PD-3D}
}

---

📄 License

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

---

📧 Contact

Ahmed Negida, MD, PhD

• 🌐 Website: https://negidamd.github.io
• 🔬 ORCID: 0000-0001-5363-6369

---

🙏 Acknowledgments

• MATLAB Deep Learning Toolbox team for the transfer learning framework
• Contributors to the original ResNet architecture

---

Advancing precision medicine in Parkinson's disease through computational neuroscience