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

Area of Science:

• Medical Imaging
• Quantum Computing
• Artificial Intelligence

Background:

• Accurate brain tumor diagnosis from MRI is challenging, especially in multi-class scenarios.
• Inter-class ambiguity and miscalibration can degrade the performance of deep learning models.
• Existing calibration methods often apply post-hoc adjustments, potentially missing opportunities for earlier refinement.

Purpose of the Study:

• To propose a novel hybrid quantum-classical framework for enhanced brain tumor diagnosis from MRI.
• To introduce a Class-wise Quantum Relational Calibration Network (CQRCNet) for improved diagnostic accuracy.
• To investigate the efficacy of quantum relational modeling for inter-class calibration in deep learning.

Main Methods:

• Developed a hybrid quantum-classical framework integrating a quantum circuit with a classical convolutional neural network.
• Employed a parameter-shared, low-qubit quantum circuit for modeling pairwise inter-class interactions.
• Integrated quantum relational modeling as an intrinsic calibration mechanism, refining logits before softmax normalization.

Main Results:

• The proposed CQRCNet demonstrated superior performance compared to classical deep learning baselines on a brain tumor MRI dataset.
• The quantum relational calibration maintained stable performance under simulated noise and finite-shot sampling conditions.
• Ablation studies confirmed the effectiveness of quantum relational modeling and provided insights into circuit design parameters.

Conclusions:

• The CQRCNet offers a promising approach for accurate and reliable multi-class brain tumor diagnosis using MRI.
• Quantum relational calibration presents a viable strategy for enhancing deep learning model performance, even in near-term quantum computing scenarios.
• The framework's robustness to noise suggests practical applicability in real-world medical imaging analysis.