An Unified Quantum Classical Model For Noisy Label Medical Image Binary Classification.

Abstract

The presence of noisy labels in medical imaging datasets can severely impact diagnostic

accuracy, leading to incorrect predictions and reduced reliability. This challenge necessitates

the development of robust classification methods capable of mitigating label noise

and ensuring consistent performance. In this study, we propose a hybrid quantum-classical

neural network (QNN-DNN) designed to enhance resilience against label noise by incorporating

quantum-assisted feature processing. The model employs quantum circuits for

feature transformation, enriching data representations before classification by a deep neural

network (DNN). By leveraging the unique properties of quantum computation, such

as entanglement and superposition, the approach effectively suppresses the adverse effects

of mislabeling. The proposed framework is evaluated on OrganMNIST and PneumoniaMNIST,

two widely used benchmark datasets in medical imaging. To systematically

assess its robustness, symmetric label noise is introduced at 10%, 20%, and 30%. Experimental

results indicate that the QNN-DNN model consistently outperforms classical

convolutional networks (CNNs) and noise-robust classification methods, demonstrating

superior accuracy under varying noise conditions. The integration of quantum feature encoding

enhances representation learning, fostering better generalization and stability despite

label inconsistencies. These findings underscore the potential of quantum-enhanced

classification frameworks in addressing label noise challenges in medical image analysis. As quantum computing technology advances, this hybrid approach could serve as a foundation for more reliable, noise-resistant AI-driven diagnostic systems, ultimately

improving patient outcomes and clinical decision-making