Premium
P4‐075: Alzheimer's disease progress estimation based on mixed‐effect biomarker modelling
Author(s) -
Guerrero Ricardo,
Schmidt-Richberg Alexander,
Ledig Christian,
Rueckert Daniel
Publication year - 2015
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.2015.06.1780
Subject(s) - biomarker , set (abstract data type) , neuroimaging , disease , computer science , cognition , artificial intelligence , estimation , machine learning , imaging biomarker , medicine , psychiatry , magnetic resonance imaging , management , economics , biochemistry , chemistry , radiology , programming language
segmentation methods were used to train a fully automated patchbased segmentation technique (Fonov et al. 2012). The method shows promising results in terms of conformity with training library (Zandifar, AAIC 2014:IC-P-150). For HarP, the 100 released HarP datasets were used (Boccardi 2015). For the Pruessner protocol, 60 subjects from ADNI data (20 NC; 20 MCI; 20 AD) were segmented by a single rater. All baseline visit ADNI1 NC and AD subjects with 1.5T scans were segmented automatically. Kappa overlap was used to evaluate segmentation accuracy in a leave-one-out validation. Cohen’s d effect size was measured between the hippocampal volume in both clinical groups in the native and stereotactic space. Results:The average Kappa overlap was 0.878 for the Pruessner protocol and 0.843 for HarP. Table 1 shows Cohen’s d effect size between the groups in both stereotaxic and native space for each side. The results show that both protocols capture the statistical difference between the group means (Cohen’s d > 0.8). The difference in effect size between segmentation methods is negligible in both native and stereotaxic space. Conclusions: Both segmentation protocols yield reasonably accurate segmentations, and have equivalent power to show NC:AD group differences. The larger kappa for Pruessner labels is due in part to tracing in 3D versus 2D for HarP labels, where boundary smoothness may not be ensured. Among the many protocols compared in (Boccardi 2014), the Pruessner protocol is the most similar to HarP, thus both methods work equally well to show group differences. Given its good performance, HarP is a good candidate to resolve heterogeneity between hippocampal segmentation protocols while maintaining power to detect volume differences between groups.