Hybrid Quantum-Classical Model for Health Risk Assessment from Skin Images

---

📋 Table of Contents

• Overview
• Background
• Technical Architecture
• Dataset Specification
• Installation
• Project Structure
• Usage
  • Training the Model
  • Making Predictions
• Output Specification
• Requirements
• Troubleshooting
• Contributing
• License

---

🔬 Overview

This project demonstrates a Hybrid Quantum-Classical Machine Learning system for assessing skin health risks. It combines classical deep learning (CNN) for feature extraction with quantum computing (Variational Quantum Classifier) for classification, creating a novel approach to non-invasive health screening.

The system analyzes skin images and classifies them into two risk categories:

• Low Risk: Healthy skin
• High Risk: Signs of vitamin C deficiency on skin

---

🎯 Background

Severe Vitamin C deficiency (scurvy) manifests on skin in various ways, including:

• Skin discoloration
• Rough, dry, scaly skin
• Corkscrew hairs
• Red or blue spots on the skin
• Poor wound healing
• Skin lesions

This project explores advanced computational models for health screening by leveraging:

1. Classical CNNs for extracting visual patterns from skin images
2. Quantum Machine Learning for enhanced classification capabilities

---

🏗️ Technical Architecture

The system implements a hybrid quantum-classical pipeline with the following workflow:

Step 1: Image Input & Preprocessing

• Accepts skin images in standard formats (JPEG, PNG, WEBP)
• Standardizes images to 224×224 pixels
• Applies normalization using ImageNet statistics

Step 2: Classical Feature Extraction

• ResNet18 backbone (pretrained on ImageNet) extracts high-level visual features
• Custom reduction layers compress 512 features → 8-dimensional feature vector
• Output features normalized to [-1, 1] range using Tanh activation

Step 3: Quantum Classification

• ZZ Feature Map encodes classical features into quantum states
• Variational Quantum Circuit with parameterized gates processes the quantum state
• Uses 8 qubits with 2 variational layers
• Entanglement pattern: full connectivity for maximum expressiveness

Step 4: Measurement & Output

• Quantum expectation values measured from circuit
• Post-processing layer maps quantum output to binary classification
• Softmax produces final risk probabilities

Input Image → CNN Feature Extractor → Quantum Classifier → Risk Assessment
  (224×224)      (512 → 8 features)      (8 qubits)        (Low/High Risk)

---

📊 Dataset Specification

The model is trained on two distinct categories:

Folder 1: healthy images skin

• Contains images of healthy skin
• Represents the negative class (Low Risk)
• ~100+ images

Folder 2: unhealthy skin

• Contains images of unhealthy skin associated with vitamin C deficiency
• Represents the positive class (High Risk)
• ~100+ images

Note: Images should show clear skin views for best results.

---

🔧 Installation

Prerequisites

• Python 3.8 or higher
• pip package manager
• 8GB+ RAM recommended
• CPU-based (quantum circuits run on CPU)

Step 1: Clone or Download the Repository

cd "Vitamin Project"

Step 2: Create a Virtual Environment (Recommended)

Windows:

python -m venv venv
venv\Scripts\activate

macOS/Linux:

python3 -m venv venv
source venv/bin/activate

Step 3: Install Dependencies

pip install -r requirements.txt

This will install:

• Qiskit 1.0+ (Quantum computing framework)
• PyTorch 2.2+ (Deep learning)
• Torchvision (Computer vision utilities)
• Pillow, OpenCV (Image processing)
• NumPy, scikit-learn (Data science)
• Matplotlib (Visualization)

Step 4: Verify Installation

python -c "import qiskit; import torch; print('Installation successful!')"

---

📁 Project Structure

Vitamin Project/
├── README.md                      # This file
├── requirements.txt               # Python dependencies
├── train.py                       # Training script
├── predict.py                     # Inference script
├── src/
│   ├── __init__.py               # Package initialization
│   ├── data_loader.py            # Data loading & preprocessing
│   ├── classical_cnn.py          # CNN feature extractor
│   ├── quantum_classifier.py     # Variational Quantum Classifier
│   └── hybrid_model.py           # Complete hybrid model
├── healthy images skin/          # Training data (healthy skin)
│   └── [100+ image files]
├── unhealthy skin/               # Training data (deficiency)
│   └── [100+ image files]
└── output/                       # Created during training
    ├── checkpoints/
    │   ├── best_model.pth        # Best model weights
    │   └── latest_model.pth      # Latest checkpoint
    ├── final_model.pth           # Final trained model
    └── training_history.png      # Training plots

---

🚀 Usage

Training the Model

Train the hybrid quantum-classical model on your skin health dataset:

Basic Training

python train.py

This uses default settings:

• Feature dimension: 8 qubits
• Quantum layers: 2
• Batch size: 4
• Epochs: 20
• Learning rate: 0.001

Advanced Training Options

python train.py \
    --healthy_dir "healthy images skin" \
    --deficiency_dir "unhealthy skin" \
    --feature_dim 8 \
    --num_layers 2 \
    --batch_size 4 \
    --epochs 20 \
    --lr 0.001 \
    --pretrained \
    --output_dir output

Key Arguments:

Argument	Description	Default
--healthy_dir	Path to healthy skin images	healthy images skin
--deficiency_dir	Path to deficiency images	unhealthy skin
--feature_dim	Number of qubits/features	8
--num_layers	Quantum variational layers	2
--batch_size	Training batch size	4
--epochs	Number of training epochs	20
--lr	Learning rate	0.001
--pretrained	Use pretrained CNN backbone	False
--test_split	Validation split ratio	0.2
--output_dir	Output directory	output

Training Output

During training, you'll see:

• Real-time progress bars with loss and accuracy
• Epoch summaries
• Best model checkpoints saved automatically
• Training history plots

Example Output:

============================================================
SKIN HEALTH QUANTUM ML TRAINING
============================================================
Loaded 100 healthy images
Loaded 100 deficiency images

Train set: 160 images
Validation set: 40 images

Quantum Circuit created with 8 qubits and 2 layers
Total parameters: 40

Epoch 1/20
Train Loss: 0.6234 | Train Acc: 68.50%
Val Loss:   0.5891 | Val Acc:   72.50%
✓ New best model saved!

---

Making Predictions

Once trained, use the model to classify new skin images:

Single Image Prediction

python predict.py "path/to/skin_image.jpg" --visualize

Example:

python predict.py "healthy images skin/normal10.png" \
    --model_path output/checkpoints/best_model.pth \
    --visualize \
    --output_dir predictions

Output:

============================================================
PREDICTION RESULTS
============================================================
Image: normal10.png
Risk Category: Low Risk
Confidence: 87.32%

Detailed Probabilities:
  Low Risk (Healthy):        87.32%
  High Risk (Deficiency):    12.68%
============================================================

Batch Prediction (Entire Directory)

python predict.py "test_images/" \
    --model_path output/checkpoints/best_model.pth \
    --visualize \
    --output_dir predictions

This will:

• Process all images in the directory
• Generate individual visualizations
• Create a CSV summary (predictions.csv)
• Display aggregate statistics

Prediction Arguments:

Argument	Description	Default
input	Image file or directory	Required
--model_path	Path to trained model	output/checkpoints/best_model.pth
--output_dir	Directory for results	None
--visualize	Generate visual plots	False

---

📤 Output Specification

The system produces one of two risk assessments:

Low Risk

• Label: "Low Risk"
• Interpretation: Healthy skin
• Class: 0
• Visual: Green indicator

High Risk

• Label: "High Risk"
• Interpretation: Signs of vitamin C deficiency on skin
• Class: 1
• Visual: Red indicator

Confidence Scores

Each prediction includes:

• Confidence: Probability of predicted class (0-100%)
• Low Risk Probability: P(Healthy)
• High Risk Probability: P(Deficiency)

---

📦 Requirements

Core Dependencies

qiskit>=1.0.2
qiskit-machine-learning>=0.7.2
qiskit-algorithms>=0.3.0
torch>=2.2.0
torchvision>=0.17.0
Pillow>=10.2.0
opencv-python>=4.9.0
numpy>=1.26.4
scikit-learn>=1.4.1
matplotlib>=3.8.3
tqdm>=4.66.2
pyyaml>=6.0.1

System Requirements

• RAM: 8GB minimum (16GB recommended for larger batches)
• CPU: Multi-core processor (quantum simulations are CPU-intensive)
• Storage: 500MB for dependencies + space for datasets
• OS: Windows 10/11, macOS 10.15+, Linux (Ubuntu 20.04+)

Note: This project runs on CPU. Quantum circuits are simulated using classical hardware.

---

🔍 Troubleshooting

Common Issues

1. Import Errors

ModuleNotFoundError: No module named 'qiskit'

Solution: Ensure virtual environment is activated and dependencies installed:

pip install -r requirements.txt

2. CUDA Warnings

CUDA not available, using CPU

Solution: This is expected and normal. Quantum circuits require CPU execution.

3. Memory Errors

RuntimeError: out of memory

Solution: Reduce batch size:

python train.py --batch_size 2

4. Slow Training

Quantum circuit simulations are computationally expensive. Solutions:

• Reduce --feature_dim (fewer qubits)
• Reduce --num_layers (simpler circuits)
• Use smaller --batch_size
• Be patient—quantum simulations take time!

5. Image Loading Errors

Error loading image: [Errno 2] No such file or directory

Solution: Check that image paths are correct and images are valid formats (JPEG, PNG, WEBP).

---

🎓 Understanding the Model

Why Quantum Machine Learning?

Quantum computing offers potential advantages:

1. High-dimensional feature spaces: Quantum states exist in exponentially large Hilbert spaces
2. Quantum entanglement: Captures complex correlations between features
3. Novel optimization landscapes: May find solutions classical models miss

Model Components

1. CNN Feature Extractor (src/classical_cnn.py)

• Pretrained ResNet18 backbone
• Extracts semantic features from images
• Reduces dimensionality to manageable size for quantum processing

2. Quantum Classifier (src/quantum_classifier.py)

• ZZ Feature Map: Encodes classical data with quantum gates
• Variational Ansatz: Trainable quantum circuit (like neural network layers)
• Measurement: Extracts classical information from quantum state

3. Hybrid Model (src/hybrid_model.py)

• Combines CNN and quantum classifier
• End-to-end differentiable (backpropagation through quantum layer)
• PyTorch-compatible for standard training loops

Training Strategy

1. Feature extractor pretraining: Use ImageNet-pretrained ResNet
2. Quantum circuit optimization: Train variational parameters via gradient descent
3. Joint fine-tuning: Optionally fine-tune entire pipeline

---

🧪 Testing the Installation

Run the individual module tests:

# Test CNN feature extractor
python -m src.classical_cnn

# Test quantum classifier
python -m src.quantum_classifier

# Test hybrid model
python -m src.hybrid_model

All tests should complete with "Test passed!" messages.

---

📈 Performance Tips

For Better Accuracy

1. Use pretrained CNN: Add --pretrained flag
2. More training epochs: Increase --epochs to 30-50
3. Data augmentation: Already included in training pipeline
4. Larger quantum circuits: Increase --num_layers (slower but more expressive)

For Faster Training

1. Smaller quantum circuits: Reduce --feature_dim and --num_layers
2. Larger batch sizes: Increase --batch_size (if memory allows)
3. Fewer epochs: Start with --epochs 10 for quick tests

---

🤝 Contributing

This is a research project demonstrating quantum machine learning concepts. Contributions, suggestions, and feedback are welcome!

Areas for Improvement

• Implement real quantum hardware backends (IBM Quantum)
• Add more sophisticated quantum feature maps
• Expand to multi-class classification
• Include explainability/interpretability tools
• Add web interface for predictions

---

📜 License

MIT License - See LICENSE file for details

---

🙏 Acknowledgments

• Qiskit: IBM's open-source quantum computing framework
• PyTorch: Facebook's deep learning library
• ResNet: Kaiming He et al.'s residual networks
• Quantum ML Community: For advancing the field

---

📚 References

1. Qiskit Machine Learning: https://qiskit.org/ecosystem/machine-learning/
2. Variational Quantum Classifiers: https://arxiv.org/abs/1804.11326
3. Quantum Feature Maps: https://arxiv.org/abs/1804.11326
4. ResNet Architecture: https://arxiv.org/abs/1512.03385

---

📞 Support

For issues, questions, or suggestions:

1. Check the Troubleshooting section
2. Review Qiskit documentation: https://qiskit.org/documentation/
3. Open an issue on the project repository

---

🎯 Success Criteria

The primary goal of this project is to demonstrate a functional, end-to-end quantum machine learning pipeline for health risk assessment. Success is measured by:

✅ Functional Pipeline: Complete workflow from image input to risk output
✅ Quantum Integration: Working variational quantum classifier
✅ Hybrid Architecture: Successful combination of classical and quantum components
✅ Binary Classification: Reliable distinction between healthy and deficiency cases
✅ Reproducibility: Clear documentation and runnable code

Note: While high accuracy is desirable, the primary focus is demonstrating the quantum computing approach as a proof-of-concept for future health screening applications.

---

🌟 Get Started Now!

# 1. Activate virtual environment
venv\Scripts\activate  # Windows
source venv/bin/activate  # macOS/Linux

# 2. Train the model
python train.py --pretrained --epochs 20

# 3. Make predictions
python predict.py "healthy images skin/normal10.png" --visualize

# 4. Explore quantum machine learning!

---

Built with quantum computing for the future of healthcare 🚀🔬