Early identification of Parkinson's disease (PD) from MRI remains challenging due to subtle structural alterations and the complexity of brain tissues. To address these challenges, this paper proposes a hierarchical framework termed Hierarchical Severity-Adaptive Diffusion Network, composed of three sequentially connected phases, where the output of each phase serves as input to the next for task-specific optimization. In the first phase, a graph diffusion-based convolutional network is employed to extract anatomical and structural features from multi-modal MRI data, enabling accurate segmentation of PD-relevant regions. Phase two introduces an edge-enhanced slice-aware recurrent network that incorporates Wiener filters and Sobel-based edge enhancement to reduce noise and partial volume effects while capturing structural continuity across adjacent MRI slices. Finally, for severity classification, non-linear severity-adaptive attention network is introduced, which emphasizes discriminative feature deterioration patterns across stages. This model uses Figshare PD dataset and demonstrates superior performance compared to established models like DenseNet121, VGG16, ResNet, MobileNet and Inception-V3, and achieves high accuracy (98.67), precision (0.99), recall (0.98), and F1 score (0.99), indicating its potential as an AI-assisted tool for PD severity assessment using MRI.