A Multimodal Adaptive Inter-Region Attention-Guided Network for Brain Tumor Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate brain tumor classification is vital for effective treatment.
• Current AI diagnostic systems face challenges with multimodal data preprocessing and feature alignment.
• Existing methods struggle to focus on shared tumor regions in 3D architectures.

Purpose of the Study:

• To develop a novel AI-based framework for advanced brain tumor classification.
• To address limitations in multimodal MRI preprocessing and cross-modal feature alignment.
• To enhance the focus on shared tumor regions within 3D neural networks.

Main Methods:

• Proposed a multimodal MRI architecture integrating Diffusion-Weighted MRI (DW-MRI) and T2-weighted MRI (T2-MRI).
• Implemented a dual-branch 3D neural architecture with a learnable High-Frequency Information Retention (HFIR) preprocessing technique.
• Utilized dual-branch 3D CNNs with an Adaptive Region Attention (ARA) module for feature extraction and alignment.

Main Results:

• The framework achieved an overall accuracy of 92.86%, sensitivity of 80.00%, and specificity of 94.12% on a brain MRI dataset.
• Statistical analyses confirmed significant outperformance compared to state-of-the-art models.
• The ARA module effectively aligned and emphasized informative shared regions across modalities.

Conclusions:

• The proposed AI framework demonstrates robust potential for precise brain tumor diagnosis.
• The novel architecture effectively overcomes limitations in multimodal data processing and feature fusion.
• This approach offers a significant advancement in AI-driven medical diagnostics for neuro-oncology.