Work

NeuroNet-X: a hybrid CNN-transformer architecture with CBAM attention for brain tumor MRI classification — multi-seed validation and explainability analysis

Year: 2026
Type: article
Abstract: Automated classification of brain tumors from MRI remains challenging due to the morphological and intensity variability across tumor subtypes. We propose NeuroNet-X, a hybrid deep learning architecture that integrates an ImageNet-pretrained ResNet-50 backbone, Convolutional Block Attention Modules (CBAM) inserted after Layer2–Layer4, a patch-embedding bridge, and a lightweight Vision Transformer (ViT) block. The framework was evaluated on three publicly available datasets: DS-1 (Bangladesh/Kaggle, 3-class), DS-2 (Figshare, 3-class), and DS-3 (Extended/Mendeley, 4-class). To address statistical reliability concerns raised in peer review, DS-2 and DS-3 results are reported as mean ± standard deviation across five independent random seeds with 95% confidence intervals. NeuroNet-X achieves $$98.34\pm 0.59\%$$ accuracy (95% CI: 97.61–99.08%) on DS-2 and $$96.66\pm 0.72\%$$ (95% CI: 95.77–97.55%) on DS-3. AUC values are $$0.9995 \pm 0.0001$$ and $$0.9929 \pm 0.0071$$ on DS-2 and DS-3 respectively. A data-leakage audit was performed prior to each experiment and duplicate training images were removed; the test sets were left intact to preserve comparability with prior work. Grad-CAM visualizations achieve a Pointing Game accuracy of 91.3% computed with manually annotated bounding boxes on DS-3 test images. These results indicate that NeuroNet-X is a technically sound decision-support framework for brain tumor categorization. However, the model was evaluated solely on retrospective 2D public benchmark datasets, and clinical deployment would require external multi-center validation, patient-level splitting, volumetric analysis, and clinician-in-the-loop reader studies.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7235665844
FWCI:
Cites: 44
Cited by:
Related to: 10