Nelissen N, Van Laere K, Thurfjell L, Buckley C, Farrar G, Brooks DJ, Vandenberghe R.
Support vector machine analysis of flutemetamol scans.
Human Amyloid Imaging 2011 Meeting Abstracts. 2011 Jan 15;
Background: Support vector machines (SVM), a supervised learning approach, can perform a strictly automated
pattern classification into one of two categories based on whole-brain images. In a purely data-driven manner, the
feature weights defined by the classifier provide information on which image components are most discriminative.
Clinically, a classifier may be useful in case of uncertainty about image interpretation.
Methods: 27 early-stage AD subjects (25 with raised uptake, eflut-pos ADf), 15 elderly controls (14 with normal
uptake, eflut-neg HVf), 10 young controls and 20 mild cognitive impairment (MCI) patients participated in the
18F-flutemetamol phase 2 study (Vandenberghe et al. Ann Neurol 2010). First, we used a leave-one-out procedure
to evaluate the diagnostic performance of an SVM with a linear kernel on the spatially normalised flut-pos AD and
flut-neg HV scans and also determined the distribution of the feature weights. The absolute distance of a subject
to the separating hyperplane normalised to unit vector (edf) gives an indication of the classificationfs robustness
Second, we evaluated performance of the algorithm in the 2 flut-neg AD subjects, 1 flut-pos HV and MCI. For
comparison, we determined how a structural MRI-based SVM (grey matter segmentation maps) categorized scans
from the 25 flut-pos AD and the 24 flut-neg HV.
Results: SVM classified all scans from flut-pos AD and from flut-neg HV correctly, with the highest absolute
feature weights in precuneus and striatum (d median 87.5, range 59.7-212.0). Automatized SVM classification of
the flut-neg AD, the flut-pos HV and the MCI was fully concordant with the visual reads. MRI-based classification
had a sensitivity and specificity of 84% and 70.8% for categorizing scans from the 25 flut-pos AD and the
24 flut-neg HV.
Conclusions: Flutemetamol scans can be reliably categorized in a strictly automated manner using an SVM.