E2E‐TM: Dual‐Way Feature Extraction and End‐to‐End Transformer Based Parkinson's Disease Diagnosis Using Integrated MR Imaging and Electroencephalogram Signals

Parkinson's disease (PD) is a liberal neurological disorder categorized by tremors, stiffness, and decreased motor function, resulting from the degeneration of dopamine-producing nerve cells in the brain. The limitations of early diagnosis of PD using ML and deep learning (DL) include potential challenges in accessing diverse and representative datasets, as well as the risk of overfitting models to specific populations, hindering the generalizability of diagnostic tools transversely diverse patient groups and demographics. To alleviate these issues, we introduced an end-to-end transformer module, E2E-TM, for precise PD diagnosis. Initially, we acquired both magnetic resonance imaging (MRI) and electroencephalography (EEG) data, underwent noise reduction using the bilateral filter and wavelet decomposition, and performed segmentation and reconstruction on MRI images using Super U-Net to reduce data complexity. Subsequently, false peaks in EEG signals were eliminated on the basis of multiple features, and both datasets were input into the proposed E2E-TM model. The transformer encoder module (TEM) included a multi-scale trunk convolution (Multi-TC) module with a penalty and reward strategy, designed in a parallel manner for feature extraction via trunk convolution. Feature maps were then mapped to their feature points using the dual-way trunk convolutional (DW-TC) module, and dual-parallel attention network (DPANet) was employed to minimize feature dimensionality. Finally, the transformer decoder module (TDM) was developed to entangle and decode the feature maps of both datasets for the classification of the diagnosed outcome. Our proposed E2E-TM model's efficiency is evaluated for proving its efficacy. As a result, our E2E-TM model attained superior diagnosis performance compared to other baseline approaches.