Brain Tumor Detection in MRI Images Using Fast Quantum-Graph Diffusion Convolutional Neural Network with Dollmaker Optimization Algorithm

Accurate and reliable brain tumor (BT) analysis from magnetic resonance imaging (MRI) remains challenging due to heterogeneous tumor appearance, fuzzy boundaries, noise sensitivity, and limited generalization of existing deep learning models across datasets. In particular, many prior works either focus solely on classification without precise segmentation, rely on fixed spatial modeling, or lack adaptive optimization, leading to reduced robustness and reproducibility. To address these gaps, this paper proposes a novel, fully integrated BT analysis framework, termed fast quantum-graph diffusion convolutional neural network with Dollmaker optimization algorithm. The proposed pipeline operates at the slice level and performs a combined task of tumor region segmentation, followed by multi-class tumor detection and classification. Specifically, MRI slices are first enhanced using BitonicX filtering for noise suppression and edge preservation, followed by fuzzy C-means with fuzzy Chebyshev distance to accurately segment irregular tumor regions. Detection and classification are then carried out using the proposed FQ-GDCNN, which integrates a quantum-inspired convolutional neural network for feature extraction with a random graph diffusion attention network to model both local and global spatial dependencies among brain regions. Notably, the term quantum refers to quantum-inspired mathematical encoding and parameterized transformations implemented on classical hardware, rather than execution on physical quantum computers. The extracted features are described as quantum-enhanced in the sense that they leverage quantum-inspired state representations and expectation-based measurements to improve feature expressiveness. Finally, the Dollmaker optimization algorithm is employed to adaptively optimize network weights, improving convergence stability and classification performance. Experimental evaluation on Figshare and BraTS 2018 datasets demonstrates that the proposed framework achieves robust and balanced performance across multiple metrics, including accuracy, precision, recall, specificity, and [Formula: see text]1-score, with a peak accuracy of 98.8%. The results confirm that the proposed method advances the state of the art by providing an adaptive, segmentation-aware, and spatially informed detection framework with improved generalization and reproducibility for automated BT diagnosis.