AMGFormer: A Multimodal Attention Graph Former Framework for Parkinson's Disease Classification Using Enhanced MRI Images

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disorder whose early diagnosis using magnetic resonance imaging (MRI) remains challenging because of subtle structural abnormalities and limited labeled medical data. Existing approaches, including radiomics-based texture analysis, multimodal MRI–genetic fusion models, and graph neural network frameworks, have demonstrated promising performance but often depend on handcrafted features, multimodal clinical data, or limited relational modeling capabilities. To address these limitations, this study proposes an Attention-guided Multimodal Graph Former (AMGFormer) for MRI-only PD classification. MRI images from the Parkinson Multi-Modal Dataset 2.0 are first preprocessed through resizing and intensity normalization and subsequently enhanced using an Adaptive Neuro-Enhancement Filter (ANEF) to improve disease-relevant anatomical structures. Multimodal representations are extracted by combining Gray Level Co-occurrence Matrix (GLCM) and Local Phase Quantization (LPQ) texture descriptors, structural morphological features, and deep semantic embeddings learned through a Medical Masked Autoencoder (MedMAE) using self-supervised learning. The fused features are transformed into a similarity graph and processed using a Graph Transformer with multi-head attention to capture complex inter-feature dependencies. The framework was implemented using Python 3.11, PyTorch 2.3, CUDA 12.1, and Scikit-Learn. Experimental results achieved 96.80% accuracy, 96.60% precision, 97.40% recall, 96.50% F1-score, and 97.00% AUC. Five fold cross-validation yielded a mean accuracy of 96.80% with a standard deviation of 0.27, demonstrating robustness and superior performance over existing MRI-based and multimodal PD diagnostic approaches.