Application and Optimization of ResNet Model Based on Spatial Attention Mechanism for Early Diagnosis of Parkinson's Disease

Magnetic resonance imaging (MRI) has been effectively proved for detecting neurodegenerative diseases such as Parkinson's disease (PD). And the application of deep learning on MRI data has become popular. This study aims to develop a deep residual network (ResNet) that incorporates a Spatial Attention Mechanism (SAM) to enhance the accuracy(ACC) of early Parkinson's disease diagnosis and classification. We utilized the structural magnetic resonance imaging (sMRI) image dataset from the Progress Marker Initiative (PPMI) database, which comprises data from 322 healthy individuals and 287 PD patients. The DICOM images were initially preprocessed, converted to NIfTI format, and aligned to MNI standard space using the FLIRT tool and the DARTEL algorithm. The SPM toolbox in MATLAB facilitated the removal of non-brain tissue and the uniform resizing of images to meet the model's input specifications. Representative median slices were then selected for model input. Our model introduces a spatial attention layer to the traditional ResNet architecture that combines maximum and average pooling to produce a spatial weight map, enhancing key features while suppressing irrelevant information. Experimental results revealed that the ResNet with spatial attention significantly outperformed the baseline model in PD classification, achieving a 96.581% ACC — an 8.547% improvement—and raising the Area Under the Curve (AUC) value varied from 0.94 to 0.97. The model's overall recognition capability has increased with almost unchanged precision, recall, and F1 score. Future work will aim to refine the attention mechanism design to boost the model's generalization capacity and practical utility. These findings are anticipated to offer new insights and technical support for the early diagnosis of neurodegenerative diseases.