Portal de Periódicos da CAPES

P2‐438: ROBUST IDENTIFICATION OF BRAIN STRUCTURES MOST DISCRIMINATIVE IN DETECTING EARLY CHANGES IN AUTOSOMAL DOMINANT ALZHEIMER'S DISEASE

2018; Wiley; Volume: 14; Issue: 7S_Part_16 Linguagem: Inglês

10.1016/j.jalz.2018.06.1130

1552-5279

Erika Molteni, Thomas Veale, André Altmann, M. Jorge Cardoso, Tammie L.S. Benzinger, Clifford R. Jack, Eric McDade, John C. Morris, Martin N. Rossor, Sébastien Ourselin, Nick C. Fox, David M. Cash, Marc Modat,

Brain Tumor Detection and Classification

Autosomal dominant Alzheimer's disease (ADAD) provides an opportunity to study presymptomatic pathophysiological changes, such as increased rates of atrophy. Very early in the disease, many years before clinical onset, pathological changes are widespread through the brain. Combining the most informative features may help track early decline. However, strong correlations between changes in brain structures could hinder the ability to reliably identify such features in multivariate analyses. Even though two highly correlated regions may be equally discriminative, some methods could consistently mark one as important and the other as irrelevant. Data comes from the tenth cutoff of the Dominantly Inherited Alzheimer Network (DIAN). This analysis included 92 mutation carriers (mean age (standard deviation) 42.7(8.5) years), 73 presymptomatic and 19 affected, and 44 non-carriers (age: 42.3(8.0) years). Volumetric T1 brain scans were parcellated into 162 unique brain structures using the Geodesic Information Flow (GIF) algorithm (Cardoso et al., IEEE Trans Med Im:34:1976-1988, 2015). Within-subject atrophy rates were calculated from two MRIs approximately 12 months apart and adjusted for total intracranial volume. To rank the most discriminative regions between ADAD mutation carriers and non-carriers, a random forest classifier was employed. Each classifier used a random subsample of 80% of the participants as the training set, from which a ranking of the features was extracted. The global ranking was obtained by averaging the rankings for each label over all the classifiers. Pairwise Pearson correlations between all structures were calculated to assess where the redundancy of information could influence the ranking. Figure. Correlation matrix showing the pairwise correlations between structures (in this image, right and left were combined for symmetrical structures). Our analysis revealed strong correlations across regional atrophy rates. These correlations must be considered when examining the most relevant features extracted by common multivariate techniques.