Parkinson's disease diagnostic support based on voxel fusion of resting BOLD signals and DTI features using multimodal pretraining.
Parkinson's disease diagnostic support based on voxel fusion of resting BOLD signals and DTI features using multimodal pretraining.
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
Beijing, CN · Author affiliation
National Engineering Research Center for Agri-Product Quality Traceability, Beijing Technology and Business University, No.11 and No.33 Fucheng Road, Haidian District, Beijing 100048, China. Electronic address: zhangqingchuan@btbu.edu.cn.Location evidence
Tongzhou, CN · Author affiliation
Business School, Beijing Wuzi University, 321 Fuhe Street, Tongzhou District, Beijing 101149, China. Electronic address: zuominbtbu@126.com.Location evidence
A plain-language reading has not been prepared for this paper yet.
Original abstract
BACKGROUND: Parkinson's disease (PD) involves concurrent changes in brain functional activity and white matter microstructure, yet single-modality analyses often fail to capture these complex alterations. NEW METHODS: We propose a voxel-level dual-stream Swin Transformer fusion framework (DSTFP) to investigate multimodal structure-function relationships in PD. DSTFP employs parallel transformer branches to extract temporal dynamics from resting-state functional MRI (rs-fMRI) and topological features from diffusion tensor imaging (DTI) fractional anisotropy maps. A cross-modal attention fusion module establishes voxel-wise correspondence between functional and structural features. RESULTS: Applied to the publicly available Parkinson's Progression Markers Initiative (PPMI) dataset, DSTFP discriminates PD, prodromal, and control groups with high robustness. Structural decoupling index (SDI) and structure-function coupling (SFC) analyses of fused features reveal distributed brain regions with characteristic alterations in functional-structural interactions. COMPARISON WITH EXISTING METHODS: DSTFP outperforms conventional single-modality and baseline multimodal models in both classification accuracy and interpretability, providing more detailed insight into voxel-level structure-function relationships. CONCLUSIONS: The proposed framework offers a robust, interpretable approach for multimodal neuroimaging analysis in PD. All source code is publicly available to support reproducibility (https://github.com/MAOmgg/DSTFP).