Alzheimer's stage progression modeling using graph neural network and MRI biomarkers

Alzheimer's stage progression modeling using graph neural network and MRI biomarkers

연구 분야: Databases

학회: Neural Computing and Applications

Stage progression and early detection are the key issues in predicting Alzheimer's disease using MRI. As of now, Alzheimer's existing work generates progression from one stage to multiple stages without the probability of progression. To overcome this problem, the graph network is very feasible for one-to-one stage progression. As for the probability of stage progression concern, this is very important to create feature vectors of biomarkers associated with the stages. The proposed work formulates the graph consisting of stages on nodes and biomarkers on edge. Each stage biomarkers are represented as a feature vector. The feature vector passes through all the stages to the destination stage and gets updated as per biomarkers. Because of many intermediate stages, more than one path is generated while traversing from one stage to another. It is also important to select the optimal path from many paths. The optimal path is selected with the inclusion of max-cut to create optimal path feature vectors. The optimal path feature vectors are multiplied by a weight matrix to produce aggregated results known as logits. After that, the logits are learned, and features are extracted during training by inputting them into a graph neural network (GNN). The proposed model adds a softmax function to GNN, which converts the trained logits into probabilities. These probabilities are scores of stage progression from one stage to another. Finally, quantum approximate optimization is used to tune the parameters and optimize the model solutions. This paper focuses on MRI for Alzheimer's early detection, one-to-one stage progression, and the probability of progression. The model achieved accuracy of 98.10% and F1 score of 97.01% for detecting AD stage. It also reported micro-precision of 99.01% and micro-recall of 99.86%. The model results illustrate the efficacy of the proposed approach.