Background: Our previous work shows that confidence in an image reader's diagnosis of Alzheimer's disease versus normal ageing is favourably influenced by superimposing the grey-scale MR images with colour overlays representing a statistical expression of each anatomical region's volume compared to a reference group. Combining recent advances in image segmentation software with high-performance computing and access to a large cohort of images of well-characterized subjects, we re-assessed the potential of this approach.
Methods: A robust and accurate automatic anatomical segmentation method, MAPER (multi-atlas propagation with enhanced registration), was applied to all T1-weighted screening MR images provided by the ADNI (Alzheimer's disease neuroimaging initiative). From each diagnostic group (HS: healthy subjects; sMCI, pMCI: mild cognitive impairment without and with progression within the observation period; AD), three test subjects were randomly chosen. A colour overlay showing regional volumetric z-score as compared to a coarsely age-matched reference group of healthy subjects was generated for each test subject and reviewed in conjunction with the grey-scale images.
Results: On visual review of the overlayed images, a typical pattern emerges. Normal subjects show predominantly green hues (z-score near zero) and few volumetric outliers. AD subjects show strongly negative z-scores for various regions predominantly located in the temporal lobe and strongly positive z-scores for ventricles, especially the temporal horn of the lateral ventricle. MCI subjects show intermediate patterns, with pMCI cases resembling the findings in AD more closely than cases of sMCI. Sample cases are shown in the figure.
Conclusions: The findings encourage us to pursue a reader study to assess classification performance. After a short training phase, readers will be shown images with regional volumetric overlays and asked to assign an individual diagnostic label.
pubs.doc.ic.ac.uk: built & maintained by Ashok Argent-Katwala.