A multi-level attention CNN-transformer based framework for the detection of brain tumor using regional dual-score...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Neuro-oncology
• Computational Neuroscience

Background:

• Deep learning models for brain tumor diagnosis often lack interpretability, limiting clinical trust and utility.
• Existing explainability methods provide only qualitative insights, failing to meet clinicians' needs for quantitative assessment.

Purpose of the Study:

• To introduce a novel classification-explainability framework for brain tumor diagnosis.
• To enhance interpretability by providing quantitative measures of explanation quality and diagnostic relevance.

Main Methods:

• Developed the Multi-Level Hybrid Network (MLHnet) integrating CNN-Transformer with multi-level attention.
• Introduced the Dual-Score Regional Explainable AI (XAI) framework for tumor region identification and geometric characteristic quantification.
• Evaluated explanation faithfulness using regional perturbation analysis.

Main Results:

• Achieved an average test accuracy of 99.30% on a brain MRI dataset of 7,023 images.
• Expert radiologist evaluation confirmed 87.64% explanation correctness.
• Dual-Score metrics successfully differentiated tumor classes based on morphological and saliency patterns.

Conclusions:

• The proposed framework offers a lightweight, high-performing, and interpretable solution for reliable brain tumor diagnosis.
• This approach addresses the limitations of conventional explainability methods in clinical settings.
• The quantitative metrics enhance diagnostic relevance and trust in AI-driven medical imaging analysis.