Dual attention based fusion network for MCI Conversion Prediction.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's disease (AD) poses a significant global health challenge.
• Early prediction of disease progression in mild cognitive impairment (MCI) is crucial for effective interventions and clinical research.

Purpose of the Study:

• To develop a novel dual attention network for classifying progressive MCI (pMCI) and stable MCI (sMCI).
• To leverage both magnetic resonance imaging (MRI) and neurocognitive metadata for enhanced classification accuracy.

Main Methods:

• Utilized a 3D CNN ShuffleNet V2 model for extracting features from MRI scans.
• Employed neurocognitive metadata to guide a spatial attention mechanism for focusing on discriminative brain regions.
• Implemented a Vision Transformer (ViT) based self-attention mechanism for fusing MRI features and neurocognitive data.

Main Results:

• Achieved an accuracy of 81.34%, AUC of 0.874, and sensitivity of 0.85 via 5-fold cross-validation.
• Demonstrated superior performance compared to existing methods for MCI progression classification.
• Ablation studies confirmed the significant contribution of both fusion methods to the model's performance.

Conclusions:

• The proposed dual attention network offers a promising approach for accurate MCI progression prediction.
• Integrating multimodal data (MRI and neurocognitive metadata) with advanced deep learning techniques enhances classification capabilities.
• This method holds potential for improving patient management and accelerating Alzheimer's disease research.