A multi-level attention CNN-transformer based framework for the detection of brain tumor using regional dual-score explainability.
Authors
Affiliations (6)
Affiliations (6)
- Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, 576104, India.
- Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, 576104, India. [email protected].
- Biolab, PolitoBIOMed Lab, Department of Electronics and Telecommunications, Politecnico di Torino, Corso Duca degli Abruzzi 24, Turin, 10129, Italy.
- Department of Medical Imaging Technology, Manipal College of Health Professions, Manipal Academy of Higher Education, Udupi, Manipal, India.
- School of Mathematics, Physics, and Computing, University of Southern Queensland, Springfield, QLD, 4300, Australia.
- Centre for Health Research, University of Southern Queensland, Springfield, Australia.
Abstract
Deep learning models for brain tumor diagnosis often lack interpretability beyond qualitative visual heatmaps. Clinicians require not only tumor localization but also quantitative assessment of explanation quality and diagnostic relevance, capabilities absent in conventional explainability methods. This paper introduces a classification-explainability framework addressing these limitations. The Multi-Level Hybrid Network (MLHnet) integrates CNN-Transformer components with multi-level attention for robust feature learning. The key novelty lies in the Dual-Score Regional XAI framework, which: (1) identifies tumor regions using hybrid saliency maps combining gradient-based and activation-based information; (2) quantifies geometric tumor characteristics via a Shape Score capturing size, circularity, and saliency concentration; and (3) evaluates explanation faithfulness using regional perturbation analysis, measuring prediction sensitivity to tumor-specific versus non-tumor regions. The framework is evaluated on a publicly available brain Magnetic Resonance Imaging (MRI) dataset containing 7,023 images across four classes. Using five‑fold cross‑validation, the proposed model achieves an average test accuracy of 99.30% (best fold: 99.47%). Expert radiologist evaluation confirms 87.64% explanation correctness, with the Dual-Score metrics successfully differentiating tumor classes based on morphological and saliency patterns. Overall, the proposed framework offers a lightweight, high-performing, and interpretable solution for reliable brain tumor diagnosis from MRI scans on the evaluated public dataset.