Quantum-Enhanced Medical Image Diagnostics

Overview

This project implements a Hybrid Quantum–Classical Model to classify images into Autistic and Non-Autistic categories.
It combines Quantum Feature Extraction using PennyLane with a Deep Neural Network (DNN) built in TensorFlow/Keras.

The goal is to demonstrate how quantum circuits can enhance feature separability in medical image analysis.

---

Technologies Used

• Quantum Framework: PennyLane (default.qubit)
• Machine Learning: TensorFlow, Keras
• Libraries: NumPy, Scikit-learn, Matplotlib, PIL
• Environment: Google Colab (Python 3.10)

---

Project Workflow

1. Dataset Loading and Preprocessing

   • Facial images (Autistic / Non-Autistic)
   • Converted to grayscale, resized (28×28), and normalized
   • Total images used: 600 (balanced classes)

2. Quantum Feature Extraction

   • 4-qubit quantum circuit
   • 2-layer RandomLayers ansatz for feature transformation
   • Generates 14×14×4 quantum feature maps

3. Hybrid Model Training

   • Deep Neural Network with:
     • Dense layers (256 → 128 → 64)
     • Batch Normalization and Dropout
     • L2 regularization
   • Optimizer: Adam (lr = 5e-5)
   • EarlyStopping for optimal convergence

4. Evaluation

   • Metrics: Accuracy, Precision, Recall, F1-score
   • Confusion Matrix for visual analysis

---

Model Architecture

Layer (Type)	Output Shape	Parameters	Description
InputLayer	(14, 14, 4)	0	Input quantum feature map (4 channels from 4-qubit circuit)
Flatten	(784)	0	Converts 2D quantum features into 1D vector
Dense (256, L2=0.001)	(256)	200,960	Fully connected layer with L2 regularization
BatchNormalization	(256)	1,024	Normalizes activations for stable training
Activation (ReLU)	(256)	0	Applies ReLU non-linearity
Dropout (0.4)	(256)	0	Randomly drops 40% neurons to reduce overfitting
Dense (128, L2=0.001)	(128)	32,896	Second fully connected layer
BatchNormalization	(128)	512	Normalizes the activations
Activation (ReLU)	(128)	0	Non-linear activation
Dropout (0.3)	(128)	0	Regularization layer
Dense (64)	(64)	8,256	Third hidden layer for deeper representation
Activation (ReLU)	(64)	0	Non-linear transformation
Dense (2, Softmax)	(2)	130	Output layer for binary classification
Total Parameters	—	243,778	Total trainable parameters

---

Training Configuration

• Optimizer: Adam (learning rate = 5e-5)
• Loss Function: Sparse Categorical Crossentropy
• Metrics: Accuracy
• Regularization: L2 (λ = 0.001) + Dropout (0.3–0.4)
• EarlyStopping: Enabled (patience = 5)
• Model Type: Hybrid Quantum–Classical Neural Network

---

Results

Metric	Value
Accuracy	70.0 %
Precision (Autistic)	0.73
Recall (Autistic)	0.66
Precision (Non-Autistic)	0.67
Recall (Non-Autistic)	0.74
Macro F1-Score	0.70

Balanced performance shows that quantum features improved the model’s ability to distinguish between both classes.

---

Visualization

• Displayed 10 random images (5×2 grid) to confirm preprocessing.
• Visualized quantum feature maps for channels 0–2 to understand feature transformation.
• Plotted training vs validation accuracy and loss.

---

Conclusion

• Increasing dataset size from 300 → 600 images improved accuracy from ~55% → 70%.
• Quantum feature extraction enhanced feature representation and model generalization.
• The hybrid approach demonstrates strong potential for quantum-assisted medical diagnostics.

---

Authors

• Abhinav Marlingaplar — 2023BCD0013
• Bhaskara Akshay Sriram — 2023BCD0015

Indian Institute of Information Technology, Kottayam
B.Tech in Computer Science (AI & Data Science)

---

License

This project is released under the MIT License.
Feel free to use or modify it for academic and research purposes.