GrAMIA-Net: Gray Matter Adaptive Multi View Imbalance Aware Attention Network for Detection of Parkinson’s Disease

Diagnosing Parkinson’s disease (PD) at its early stages from structural MRI remains a difficult task, primarily because of scanner-dependent variations, skewed class distributions, and the subtle anatomical alterations associated with early neurodegeneration. To improve MRI-based Parkinson’s disease classification under these conditions, GrAMIA-Net is proposed as a multi-view deep learning framework built upon gray matter representations derived from T۱-weighted MRI. The framework learns complementary information from sagittal, coronal, and axial perspectives and employs a hierarchical attention mechanism to emphasize discriminative spatial cues within each view. In addition, an inter-view attention module is utilized to aggregate information across different anatomical planes effectively. To mitigate the adverse effects of class imbalance, an Adaptive Imbalance Optimization (AIBO) scheme is integrated into the training process, thereby improving learning for underrepresented samples. Validation on the multi-scanner and class-imbalanced PPMI cohort shows that the proposed framework attains an accuracy of ۰.۸۶۲, an F۱-score of ۰.۹۱۹, and an AUC of ۰.۷۲۰, demonstrating reliable predictive capability across categories. These results suggest that GrAMIA-Net provides a practical and computationally efficient framework for PD detection using single-modality GM representations, with potential applicability in real-world neuroimaging settings.

کلیدواژه ها:

Adaptive Imbalance Optimization ، Gray Matter Maps ، Multi-View Learning ، PD Detection ، T۱-weighted Structural Imaging