MoRE-Net: An Interpretable and Modality-robust Model for Brain Tumor Grading.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Machine Learning for Diagnostics

Background:

• Trustworthy artificial intelligence (AI) requires both interpretability and robustness, especially in critical medical diagnostic applications.
• Existing interpretable AI models often lack robustness, particularly when dealing with incomplete or missing data modalities.
• Improving the resilience of interpretable diagnostic models to missing data is crucial for clinical adoption.

Purpose of the Study:

• To enhance the robustness of interpretable multimodal medical imaging diagnostic models under missing modality conditions.
• To develop a novel AI framework that maintains diagnostic accuracy and interpretability despite data gaps.
• To address the challenge of inter-modality interaction absence in multimodal diagnostic AI.

Main Methods:

• Propose the Modality-Robust and Explainable Network (MoRE-Net), utilizing per-modality encoders and a Mamba architecture for efficient global-context modeling.
• Introduce an online multimodal teacher to guide per-modality encoders via alignment loss during early training stages.
• Evaluate MoRE-Net on the BraTS2020 and ReMIND datasets for brain tumor grading, assessing performance with balanced accuracy (BAC) and interpretability with activation precision (AP).

Main Results:

• MoRE-Net achieved an average balanced accuracy (BAC) of 73.5% and activation precision (AP) of 61.2% across missing modality scenarios on the BraTS2020 dataset.
• The model outperformed baseline methods by approximately 15% in BAC and 21% in AP.
• Validation on the ReMIND dataset and ablation studies confirmed the effectiveness of individual strategies and the overall robustness of MoRE-Net.

Conclusions:

• MoRE-Net is a novel interpretable and modality-robust AI model for brain tumor grading.
• The model demonstrates significant improvements in diagnostic accuracy and interpretability, even with missing data.
• MoRE-Net shows considerable potential for reliable clinical deployment in medical diagnostics.