Mining fMRI Dynamics with Parcellation Prior for Brain Disease Diagnosis

Xiaozhao Liu,Mianxin Liu,Lang Mei,Yuyao Zhang,Feng Shi,Han Zhang,Dinggang Shen
2023-05-04
Abstract:To characterize atypical brain dynamics under diseases, prevalent studies investigate functional magnetic resonance imaging (fMRI). However, most of the existing analyses compress rich spatial-temporal information as the brain functional networks (BFNs) and directly investigate the whole-brain network without neurological priors about functional subnetworks. We thus propose a novel graph learning framework to mine fMRI signals with topological priors from brain parcellation for disease diagnosis. Specifically, we 1) detect diagnosis-related temporal features using a "Transformer" for a higher-level BFN construction, and process it with a following graph convolutional network, and 2) apply an attention-based multiple instance learning strategy to emphasize the disease-affected subnetworks to further enhance the diagnosis performance and interpretability. Experiments demonstrate higher effectiveness of our method than compared methods in the diagnosis of early mild cognitive impairment. More importantly, our method is capable of localizing crucial brain subnetworks during the diagnosis, providing insights into the pathogenic source of mild cognitive impairment.
Image and Video Processing,Neurons and Cognition
What problem does this paper attempt to address?
The paper attempts to address the issue of how to more effectively utilize spatiotemporal information in functional magnetic resonance imaging (fMRI) data for brain disease diagnosis. Specifically, existing methods often fail to fully leverage the temporal fluctuations related to disease diagnosis and overlook the importance of specific subnetworks within brain functional networks (BFNs). Therefore, the authors propose a novel graph learning framework—SLMIL-GCTrans, aiming to improve diagnostic performance through the following two main innovations: 1. **Spatiotemporal Feature Extraction**: Combining Graph Convolutional Networks (GCN) and Transformer models (named the GCTrans module) to more efficiently extract spatiotemporal features. This approach can better capture the dynamic changes in fMRI data, thereby constructing higher-level brain functional networks. 2. **Subnetwork Prioritization**: Utilizing established brain segmentation knowledge and applying an attention mechanism-based multiple instance learning strategy (SLMIL) to highlight key subnetworks affected by the disease. This strategy not only improves diagnostic performance but also enhances the interpretability of the model. Experimental results show that this method outperforms other methods in the diagnosis of early mild cognitive impairment (EMCI) and can effectively locate key brain regions leading to mild cognitive impairment, providing valuable insights into understanding its pathological origins.