A class-wise quantum relational calibration network for brain tumor diagnosis

Accurate and reliable brain tumor diagnosis from magnetic resonance imaging (MRI) remains a challenging task, particularly in multi-class settings where inter-class ambiguity and miscalibration may degrade predictive performance. In this work, we propose a Class-wise Quantum Relational Calibration Network (CQRCNet), a hybrid quantum-classical framework that integrates quantum relational modeling into a classical convolutional neural network backbone. Our approach employs a parameter-shared and low-qubit quantum circuit to explicitly model pairwise interactions between class-specific evidences, serving as a calibration mechanism for logit refinement. Unlike conventional post-hoc calibration methods, the proposed mechanism operates by refining inter-class relational representations prior to softmax normalization. Empirical studies are conducted on a publicly available brain tumor MRI dataset covering four tumor categories, and the proposed method demonstrates favorable performance compared with several reported classical deep learning baselines and representative hybrid quantum-classical models. Noise-aware quantum simulations under depolarizing noise and finite-shot sampling further indicate that the proposed quantum relational calibration maintains stable performance under simulated realistic quantum conditions, suggesting its potential applicability to near-term quantum computing scenarios. In addition, ablation studies, including comparisons with temperature scaling and analyses of different quantum circuit depths, provide further insights into the role of quantum relational modeling in the proposed framework.