Author: Ji Hye Han, Yookyung Kim, Jang-Hoon Oh, Heesoon Sheen, Han-Back Shin 👨🔬
Affiliation: Ewha Womans university, Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, High-Energy Physics Center, Chung-Ang Universit, Ewha Womans University, Kyung Hee University Hospital 🌍
Purpose: Chest X-rays are critical for diagnosing conditions such as pneumonia, tuberculosis, and COVID-19. Although deep learning (DL) approaches, especially convolutional neural networks, have significantly advanced image classification, their generalizability across diverse datasets remains limited. Quantum deep learning (QDL) harnesses quantum phenomena like superposition and entanglement to potentially address these limitations. This study evaluates a hybrid quantum-classical convolutional neural network (QCCNN) compared to a traditional ResNet50 model, aiming to enhance diagnostic efficiency and accuracy through quantum computing’s high-dimensional data processing.
Methods: The study implemented and compared a conventional ResNet50 model with a hybrid quantum-classical model incorporating quantum circuits. The hybrid model featured a quantum synthesis layer integrated into the classical neural network architecture to enhance learning efficiency. Implementation utilized the Pennylane software platform, accessing quantum operations through IBM Quantum and IonQ quantum hardware API tokens. Both models were evaluated on their ability to classify various chest pathologies, including COVID-19, tuberculosis, and pneumonia.
Results: Overall, the classical ResNet50 model achieved higher accuracy, precision, and efficiency in training and evaluation. However, the QCCNN demonstrated competitive performance in detecting specific pathologies, notably COVID-19 and tuberculosis. These results highlight strengths in targeted disease detection, although the hybrid approach did not surpass the classical model on all metrics.
Conclusion: Quantum deep learning shows potential in medical imaging. While the hybrid QCCNN did not universally outperform the classical ResNet50, its comparable performance for certain classes indicates a promising future. Advancements in quantum hardware and further optimization of hybrid architectures are essential for achieving parity with, and possibly exceeding, classical deep learning methods in diagnostic imaging.