Hybrid Quantum-Classical CNN for Medical Image Diagnosis

This repository presents a Hybrid Quantum-Classical Convolutional Neural Network (Q-CNN) for automated finding of chest diseases from medical images. By combining the feature extraction strength of classical deep learning with the high-dimensional expressivity of Quantum Machine Learning (QML), the project demonstrates a scalable pipeline for real-world datasets such as CheXpert.

---

🎯 Core Objectives

• Integrate PennyLane quantum computing with high-performance JAX / PyTorch workflows.
• Build an efficient data pipeline for large-scale medical imaging datasets.
• Design a hybrid architecture that improves diagnostic performance while reducing computational overhead.

---

🧠 Technical Architecture

The system follows a Parallel Residual Hybrid Design to stabilise training and mitigate the Barren Plateau issue common in quantum circuits.

1) Classical Feature Extraction

• Uses a pre-trained ResNet18 to extract 512-dimensional feature vectors from 224×224 X-ray images.
• Pre-computes image features to decouple heavy processing.
• Accelerates hybrid training by ~50×.

2) Quantum Path — “Booster”

• Compresses classical features into 10 qubits using dense layers.
• Applies AngleEmbedding and StronglyEntanglingLayers.
• Explores correlations in a 2¹⁰ Hilbert space to enhance classification.

3) Classical Path — “Safety Net”

• Runs a parallel dense network to match baseline neural performance.
• Merges outputs using a residual connection:

[ \text{Final Output} = \text{Classical Output} + \text{Quantum Output} ]

---

⚙️ Key Features & Techniques

• High-Performance Computing: JAX JIT compilation with XLA optimization for faster quantum simulation.
• Uncertainty Handling: Recall optimization by mapping uncertain labels to potentially diseased states.
• Class Imbalance Handling: Focal Loss (γ = 2.0) to prioritize hard minority classes.
• Clinical Tuning: Dynamic decision thresholding to maximize F1-Score per disease risk.

---

📊 Performance Results

Metric	Result
Best Validation Accuracy	86.85%
Macro AUROC	0.7035
Recall (Atelectasis)	91%
Recall (Edema)	95%

Note: Designed as a high-sensitivity first-pass screening model, prioritizing Recall to minimize missed pathologies.

---

🚀 Future Scope

• Scale to more qubits and deeper quantum layers.
• Deploy on real Quantum Processing Units (QPUs).
• Extend to 3D CT scan data processing.

---

Datset

• Dataset available at https://www.kaggle.com/datasets/ashery/chexpert