Premium
IC‐P‐057: CLASSIFICATION OF PATHOLOGY USING BRAIN SUBSTRUCTURE VOLUMES IN POST MORTEM CONFIRMED DEMENTIAS
Author(s) -
Harper Lorna,
Burton Emma,
Bouwman Femke H.,
Rozemuller Annemieke,
Barkhof Frederik,
Scheltens Philip,
O'Brien John,
Fox Nick,
Ridgway Gerard R.,
Schott Jonathan M.
Publication year - 2014
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1016/j.jalz.2014.05.062
Subject(s) - dementia , frontotemporal lobar degeneration , medicine , frontotemporal dementia , feature selection , pathology , dementia with lewy bodies , artificial intelligence , pattern recognition (psychology) , disease , computer science
Background:Classification of dementia based onmeasures of brain volume has been studied previously. However, many of these studies use clinical diagnosis as the gold standard and/or investigate group separation of particular dementias from healthy controls. In this study, we investigate the more clinically important problem of differential diagnosis of cases with pathologically confirmed Alzheimer’s disease (AD), dementia with Lewy bodies (DLB), frontotemporal lobar degeneration (FTLD) or other pathology. Using retrospective data from patients who attended clinics in three different European centres (VUMC-Amsterdam, DRC-London, IAH-Newcastle) we evaluated the diagnostic utility of applying automated segmentation and machine learning to this clinically realistic dataset.Methods:Multi-label segmentation propagation, based on 83 manually segmented regions from 30 subjects, was applied to T1-weighted volumetric MRI scans from 305 patients with pathologically-confirmed dementia (127 AD, 49 DLB, 80 FTLD, 49 other). Segmented volumes were corrected for age, gender, total intracranial volume and scanner, then scaled to zero mean and unit variance over subjects. The dataset was then divided into training (n1⁄4153) and testing sets (n1⁄4152), preserving pathology proportions. Univariate feature selection was applied such that only the top 10% of regions most important for prediction were used in the models. Group separation was quantified using a linear support vector machine with cross-validation. Class weighting was applied to adjust for unbalanced groups. Three classifiers were trained to separate each of the three main dementia groups from the pooled group of all other dementias. A further three classifiers were trained to discriminate each pair of the main dementia groups. Results: Feature selection demonstrated higher accuracies for group separation than using all features. Central structures including the putamen, thalamus and substantia nigra, and temporal lobe regions helped to distinguish FTLD from DLB. AD and FTLD were best distinguished using frontal/occipital lobe regions, and AD and DLB using temporal/parietal regions (Table1). Classification accuracy was greatest for AD pathology in all models (1). Conclusions: Classification of dementia pathology based on brain substructure volumes may help to maximise the diagnostic information available from clinical