This poster is published under an
open license. Please read the
disclaimer for further details.
Keywords:
Neuroradiology brain, Computer applications, MR, Computer Applications-Detection, diagnosis, Dementia, Outcomes
Authors:
L. Sørensen1, M. Lillholm1, A. Pai1, I. Balas1, C. Anker2, C. Igel1, M. Nielsen1; 1Copenhagen/DK, 2Kgs. Lyngby/DK
DOI:
10.1594/ecr2015/B-0077
Methods and materials
Two T1-weighted structural MRI reference datasets were considered.
ADNI: baseline scans from the "complete annual year 2 visit" 1.5T standardized Alzheimer’s Disease Neuroimaging Initiative dataset [2] (169 normal controls (CTRL),
234 subjects with mild cognitive impairment (MCI),
101 AD patients).
AIBL: baseline scans from the imaging arm of the Australian Imaging,
Biomarker & Lifestyle Flagship Study of Aging [3] (88 CTRL,
29 MCI,
and 28 AD).
The challenge-winning combination MRI biomarker was applied to each scan.
First, the following individual MRI biomarkers were computed:
- cortical thickness using cross-sectional FreeSurfer,
- hippocampal shape using an in-house method,
- hippocampal texture using an in-house method,
- standard volumetry using cross-sectional FreeSurfer as well as an in-house method specifically constructed to segment the hippocampus.
The individual MRI biomarkers were subsequently age-normalized and combined using a regularized linear discriminant analysis classifier (LDA).
We refer to [4] for further details.
The method was applied using 10-fold cross-validation stratified on diagnostic group and cohort,
and performance on ADNI and AIBL was subsequently investigated separately for the two datasets.
Both per-class and three-class receiver operating characteristic (ROC) curves and the associated area under the ROC curve (AUC) were computed [1].
The evaluation python scripts supplied by the CADDementia organizers were used for this purpose.