Comparison of automated brain segmentation using a brain phantom and patients with early Alzheimer's dementia or mild cognitive impairment

Magnetic resonance imaging (MRI) and brain volumetry allow for the quantification of changes in brain volume using automatic algorithms which are widely used in both, clinical and scientific studies. However, studies comparing the reliability of these programmes are scarce and mainly involved MRI derived from younger healthy controls. This study evaluates the reliability of frequently used segmentation programmes (SPM, FreeSurfer, FSL) using a realistic digital brain phantom and MRI brain acquisitions from patients with manifest Alzheimer's disease (AD, n=34), mild cognitive impairment (MCI, n=60), and healthy subjects (n=32) matched for age and sex. Analysis of the brain phantom dataset demonstrated that SPM, FSL and FreeSurfer underestimate grey matter and overestimate white matter volumes with increasing noise. FreeSurfer calculated overall smaller brain volumes with increasing noise. Image inhomogeneity had only minor, non- significant effects on the results obtained with SPM and FreeSurfer 5.1, but had effects on the FSL results (increased white matter volumes with decreased grey matter volumes). The analysis of the patient data yielded decreasing volumes of grey and white matter with progression of brain atrophy independent of the method used. FreeSurfer calculated the largest grey matter and the smallest white matter volumes. FSL calculated the smallest grey matter volumes; SPM the largest white matter volumes.Best results are obtained with good image quality. With poor image quality, especially noise, SPM provides the best segmentation results. An optimised template for segmentation had no significant effect on segmentation results. While our findings underline the applicability of the programmes investigated, SPM may be the programme of choice when MRIs with limited image quality or brain images of elderly should be analysed.