Multi-Modal Diagnosis of Alzheimer's Disease Using Interpretable Graph Convolutional Networks

Abstract

The interconnection between brain regions in neurological disease encodes vital information for the advancement of biomarkers and diagnostics. Although graph convolutional networks are widely applied for discovering brain connection patterns that point to disease conditions, the potential of connection patterns that arise from multiple imaging modalities has yet to be fully realized. In this paper, we propose a multi-modal sparse interpretable GCN framework (SGCN) for the detection of Alzheimer's disease (AD) and its prodromal stage, known as mild cognitive impairment (MCI). In our experimentation, SGCN learned the sparse regional importance probability to find signature regions of interest (ROIs), and the connective importance probability to reveal disease-specific brain network connections. We evaluated SGCN on the Alzheimer's Disease Neuroimaging Initiative database with multi-modal brain images and demonstrated that the ROI features learned by SGCN were effective for enhancing AD status identification. The identified abnormalities were significantly correlated with AD-related clinical symptoms. We further interpreted the identified brain dysfunctions at the level of large-scale neural systems and sex-related connectivity abnormalities in AD/MCI. The salient ROIs and the prominent brain connectivity abnormalities interpreted by SGCN are considerably important for developing novel biomarkers. These findings contribute to a better understanding of the network-based disorder via multi-modal diagnosis and offer the potential for precision diagnostics. The source code is available at https://github.com/Houliang-Zhou/SGCN.

Fig. 1.

An overview of the proposed SGCN model for Alzheimer’s diagnosis and biomarker interpretation. The multi-modal brain images are converted to graphs by using the Gaussian similarity to construct the connections between ROIs. The graphs combined with the feature importance probability ${\mathit{P}}_{\mathit{X}}$ and the edge importance probability ${\mathit{P}}_{\mathit{A}}$ are sent to our proposed sparse GCN model to predict the disease. The importance probabilities ${\mathit{P}}_{\mathit{X}}$ on nodes and ${\mathit{P}}_{\mathit{A}}$ on edges provide the interpretation for the salient ROIs and the prominent disease-specific connections.

Fig. 2.

Multiclass classification comparison between the state-of-the-art machine learning models and our proposed SGCN by using different modalities. The highest a) accuracy, b) sensitivity, and c) specificity labeled with a white star are 0.826, 0.804, and 0.845 respectively, which are achieved by our SCGN when using all three modalities.

Fig. 3.

The interpretation of salient ROIs and the most discriminative brain connections in distinguishing AD from HC. a) Interpreting top 20 salient ROIs based on the importance probability ${\mathit{P}}_{\mathit{X}}$ between different modalities. The commonly detected salient ROIs across different modalities are circled in blue. b) Comparison between the KNN graph and the sparse interpretation of prominent brain network connections in AD group. c) The significant difference of the interpreted most discriminative connections for distinguishing HC and AD was evaluated by two-sample t-tests with false discovery rate (FDR) corrected p-value < 0.05. Here, the top 20 most discriminative ROI connections are visualized for interpretation by using multi-modalities. The dark-red and dark-blue color indicates the high positive and low negative t values.

Fig. 4.

The interpretation of salient ROIs and the most discriminative brain connections in distinguishing MCI from HC. This interpretation in MCI was reported by using the same strategy from AD analysis.

Fig. 5.

Prediction of ADAS13, MMSE, and CDR-SOB test scores using multiple linear regression based on the ROI features learned by the last GCN layer. The prediction performance was evaluated using 5-fold cross-validation. The significance of the prediction was confirmed by random permutation tests of 10000 times. The actual correlation coefficients between the predicted scores and true scores are indicated by red dashed lines.

Fig. 6.

Neural system-level interpretation of the most discriminative connections. a) The absolute t value of most discriminative ROI connections with FDR correlated p-value < 0.05 were reported between neural systems by using multi-modalities. The dark-red color indicates a high score. The non-significant connections are marked as white. b) Such t values were reported by using different modalities in each neural system. Here, the reported $\mathit{t}$ values of one modality were the average results over all single modalities. Similarly, the $\mathit{t}$ values of two modalities were the average results over all three pairs of modalities.

Fig. 7.

Interpreting top 20 salient ROIs between males and females under multi-modalities.