Graph Morphology-Based Genetic Algorithm for Classifying Late Dementia States

Abstract

Early diagnosis of neurological diseases such as Alzheimer’s disease (AD) is extremely vital for patient treatment. Analyzing the human brain connectivity is a popular approach in investigating the relationship between the brain morphology, structure, and function and the emergence of neurological diseases. However, extracting relevant diagnostic information from the connectome is still one of the most challenging problems. Many works have thoroughly studied the connectional map of the brain, however, to the best of our knowledge, no previous study had used graph morphology to rigorously explore the topological properties of the human connectome. In this paper, we propose a novel graph morphology-based genetic algorithm (GMGA) to mine the brain network and extract the most relevant connections for disordered brain state stratification. First, we define our graph morphological structural operators (SE) and design a subgraph matching technique for matching a particular graph-based SE with an input brain connectome. Second, we propose GMGA which identifies the optimal sequence of morphological operations using a predefined structural element for distinguishing between two brain states (e.g., late mild cognitive impairment (LMCI) vs Alzheimer’s disease (AD)). Last, we train a linear classifier in a K-fold cross-validation fashion using the morphed brain graphs given the optimal learned morphological operator sequence. Our experimental results demonstrate a significant gain in classification performance between LMCI and AD groups in comparison with baseline methods. This work constitutes the first proof-of-concept of the merit of graph morphology in decoding the healthy and disorder brain connectomes.