IC‐P‐073: DISSIMILARITY BASED EXTRACTION OF COVARIANCE LINKED NETWORK (DECLINE) FEATURES FOR EARLY DETECTION OF AD

Cortical thickness analysis is a powerful tool to assess the onset and progression of neurodegenerative diseases such as Alzheimer's. Spatial gradients in cortical thinning are a hallmark of dementias, and are shown to follow stereotypical patterns specific to each dementia. There is tremendous variability of cortical thickness across the population, but the signature of the disease is much more clearly visible in cortical thickness gradients taken between different brain regions, for example anterior-posterior gradients in AD as AD is known to affect the posterior cortices such as the medial temporal lobes, the precuneus, parietal areas, entorhinal cortex preferentially and early in the course of the disease. Keeping this idea of preferential gradient is mind, We reformulate our previously developed novel imaging cortical biomarker based on graph-theoretic analysis of inter-regional covariance of cortical thickness, called ThickNet features [1] to develop novel covariance features based on dissimilarity in thickness. These features capture the spatial thickness gradients within each subject. We call them Dissimilarity based Extraction of Covariance LInked NEtwork (DECLINE) features. We show that they outperform ThickNet features for the early detection of Alzheimer's disease. DECLINE features are first of its kind and show promise in detecting the cognitive decline predictive of AD. Cortical thickness is extracted from the MRI scan using the method [2] for each patient and the features are registered to a common atlas surface to establish vertex-wise correspondence. Then the cortex of each patient is partitioned into large number (e.g. K=300 vertices per patch) of small areas by k-means clustering of vertices spatially on the atlas surface (Figure 1). A graph is then constructed by establishing a link between two such patches if dissimilarity in thickness is above a given threshold (e.g. 0.5mm). Dissimilarity is defined as abs. difference in mean-thickness from the two patches. From this binary undirected graph, we compute several graph-theoretic properties - called DECLINE features - to represent each patient (See Figure 2), namely nodal degree, betweenness centrality and clustering coefficient. DECLINE features are intrinsic to each patient and offer a novel insight into neurodegeneration caused by various diseases. Using the same dataset (ADNI) on which we demonstrated the diagnostic utility of ThickNet features, we show that DECLINE features outperform ThickNet features, and show potential for the early detection of Alzheimer's disease. Using our Repeated Holdout, Stratified Training set (RHsT) cross-validation method proposed in [1] (See Figure 3), We evaluate the performance of DECLINE features. They produced an area under ROC (AUC) of 0.93 in discriminating AD from healthy controls (CN), and an AUC of 0.87 in discriminating MCI converters (MCIc) from CN. We also present results for two other experiments: MCIc vs. MCI non-converters (MCInc) and CN vs. MCI (MCIc+MCInc). The results presented in Figure 4 show DECLINE features significantly outperform ThickNet features in all the experiments. We present novel DECLINE measures based on inter-regional covariance of cortical thickness from structural MRI. We demonstrate their diagnostic utility in four classification experiments on benchmark ADNI dataset, and esp. for early detection of Alzheimer's disease. DECLINE features show promise for predictive modeling in neuroimaging applications. Visualization of Cortical Parcellation Resulting from our Method, with approx. 300 vertices per patch. DECLINE Feature extraction. cross-validation to evaluate classification performance of the features. Comparison of Best Classification Results in each binary experiment for the two methods. References: 1. Raamana, PR, Wang, L, Beg, MF (2013). Thickness NETwork (ThickNet) Features for the Detection of Prodromal AD. Machine Learning in Medical Imaging, 8184 (Chapter 15), 114-122. http://dx.doi.org/10.1007/978-3-319-02267-3_15. 2. Gibson E, Wang L, Beg MF. Cortical thickness measurement using Eulerian PDEs and surface-based global topological information. Org Human Brain Mapping, 15th Annual Meeting. 2009.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.