Explainable Brain Tumor Classification Using a Hybrid Deep Learning Framework with Quantum-Inspired Feature Enhancement and Adaptive Feature Selection

Abstract

Accurate and interpretable classification of brain tumors from magnetic resonance imaging (MRI) is essential for timely diagnosis and treatment planning. Despite substantial progress in deep learning-based medical image analysis, existing models often suffer from limited feature discrimination, poor generalization across heterogeneous tumor appearances, and a lack of clinically meaningful explainability. In this paper, we propose a hybrid deep learning framework that integrates a convolutional neural network backbone with a lightweight vision transformer branch, augmented by a quantum-inspired feature enhancement (QIFE) module and an adaptive feature selection (AFS) mechanism. The QIFE module leverages a parameterized quantum-inspired transformation, simulating unitary operations on feature embeddings to amplify salient patterns and suppress noise without the need for actual quantum hardware. The AFS component dynamically reweights channel-wise feature responses through a gated attention structure, enabling the model to focus on the most diagnostically relevant regions. The proposed architecture was evaluated on a publicly available multi-class brain tumor MRI dataset comprising 3064 contrast-enhanced T1-weighted images across glioma, meningioma, and pituitary tumor categories. Extensive experiments demonstrate that the hybrid framework achieves an overall classification accuracy of 97.83%, with precision, recall, and F1-score all exceeding 97.5%, and an area under the receiver operating characteristic curve above 0.996. Ablation studies confirm that both QIFE and AFS contribute significantly to performance gains, yielding an improvement of 2.4% in accuracy over a comparable baseline without these components. Explainability analyses using gradient-weighted class activation mapping and attention map visualization reveal that the model consistently focuses on pathologically meaningful tumor boundaries and internal structures, aligning with radiological expertise. Statistical tests further validate the significance of the improvements. The proposed approach offers a promising direction for building high-performance, interpretable computer-aided diagnosis systems, combining the representational power of hybrid architectures with biologically inspired and quantum-inspired mechanisms for enhanced feature refinement.