Hybrid Classical-Quantum Transfer Learning for Cardiomegaly Detection in Chest X-rays

Area of Science:

• Quantum computing applications in healthcare
• Machine learning for medical image analysis
• Cardiovascular disease diagnostics

Background:

• Cardiovascular diseases pose significant health challenges.
• Chest X-rays (CXRs) are crucial for diagnosing conditions like cardiomegaly.
• Quantum machine learning (QML) offers potential advancements in medical imaging analysis.

Purpose of the Study:

• To develop and evaluate hybrid classical-quantum (CQ) transfer learning models for detecting cardiomegaly in CXRs.
• To compare the performance of CQ models against classical-classical (CC) models.
• To assess the interpretability of CQ models using heatmaps.

Main Methods:

• Designed hybrid CQ transfer learning models integrating parameterized quantum circuits (using Qiskit and PennyLane) with classical networks (PyTorch).
• Utilized a balanced dataset of 2436 posteroanterior CXRs from the CheXpert repository.
• Employed k-fold cross-validation and a state vector simulator for training CQ models.
• Analyzed trainability using normalized global effective dimension.

Main Results:

• Achieved high predictive performance with ROC AUC scores up to 0.93 and accuracies up to 0.87, comparable to CC models.
• Demonstrated significantly more frequent visualization of trustworthy Grad-CAM++ heatmaps covering the heart with CQ models (94%) compared to CC models (61%).

Conclusions:

• Hybrid classical-quantum models show strong potential for accurate cardiomegaly detection in CXRs.
• The enhanced interpretability of CQ models may increase their adoption by healthcare professionals.
• QML represents a promising frontier for improving cardiovascular diagnostics through medical imaging.