Brain tumors represent one of the deadliest types of neurological diseases, a fact that makes early accurate diagnosis essential for the survival of patients. Tumor analysis based on deep learning, as it is currently done, has three main problems: the segmentation step fails to adequately model the long-range spatial dependencies, the classification pipelines have the problem of feature redundancy, and there is no modeling of the structural relationship between the featuresall of these factors together make the diagnosis less reliable. This work proposes a novel Quantum-Inspired Optimization of Transformer-Capsule Networks (QI-TCN) framework that jointly addresses these limitations. The framework integrates: (i) a Swin Transformer U-Net (Swin-UNet) for hierarchical segmentation; (ii) Quantum-Inspired Harris Hawks Optimization (QHHO) for both segmentation hyperparameter tuning and discriminative feature selection; (iii) EfficientNetV2 for compact high-capacity feature extraction; and (iv) a Graph Attention Capsule Network (GACN) for spatially-aware tumor classification. Evaluated on the BraTS 2019 and BraTS 2020 benchmarks, the proposed framework achieves classification accuracies of 99.02% and 99.15%, F1-scores of 98.97% and 99.10%, Dice-WT scores of 97.4% and 97.4%, and AUC values of 99.3% and 99.4%, respectivelysurpassing all compared baselines. All improvements are statistically significant (paired t-test, t = 12.84, p < 0.001) across 5 independent runs. Cross-dataset evaluation on the Figshare Brain Tumor Dataset yields 96.8% zero-shot accuracy and 98.4% after 10-epoch fine-tuning. These results demonstrate that the proposed framework offers a robust, interpretable, and clinically viable solution for automated brain tumor diagnosis.
Aarti et al. (Thu,) studied this question.