MEST: Multi-plane Embedding and Spatial-temporal Transformer for Parkinson's disease diagnosis

• Published in: 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) pp. 1072 - 1077
• Main Authors: Liu, Jin, Du, Hao, Bi, Qian, Liao, Haiyan, Pan, Yi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06.12.2022
• Subjects: Ensemble learning; Functional magnetic resonance imaging; Medical diagnosis; Multi-modality; Multiplane; Parkinson's disease; Redundancy; Representation learning; Spatial-temporal; Time series analysis; Transformers
• DOI: 10.1109/BIBM55620.2022.9995498

Parkinson's disease (PD) is a common neurodegenerative disorder that impairs the patient's quality of life. Medical imaging technology has provided a variety of neuroimages for PD diagnosis studies. However, how to effectively integrate the rich representations from multi-modality data is still a challenging task. To address this challenging task, we propose a multiplane embedding and spatial-temporal Transformer (MEST) framework for PD diagnosis. Firstly, we project structural magnetic resonance imaging (sMRI) into 2D images from coronal, sagittal and axial directions, respectively. Then, the multi-plane dynamic images are learned by pre-trained VGG11 and attention mechanism for representation learning. Afterwards, time series information of functional magnetic resonance imaging (fMRI) is used to construct dynamic functional connection images. To capture information changes in brain, spatial-temporal connectivity Transformer (SCTransformer) is utilized to address spatial-temporal redundancy and dependencies. To integrate multimodality data, ensemble learning is designed based on majority voting strategy to perform PD diagnosis. We evaluate the proposed method on 279 subjects from an in-house and Parkinsons Progression Markers Initiative (PPMI) dataset. Experimental results show that the MEST achieves promising performance with accuracies of 0.856 and 0.885, and outperforms some state-of-the-art methods.