A comparison of methods for the automated calculation of volumes and atrophy rates in the hippocampus

Abstract

Hippocampal atrophy rates have been used in a number of studies in Alzheimer’s disease (AD) to assess disease progression and are being increasingly utilized as an outcome measure in clinical trials of new pharmaceutical agents. Owing to the labor-intensive nature of hippocampal segmentation, more automated approaches are required for such analysis. In this study we compared methods of automatically segmenting the hippocampus (single-person template and template library) on the baseline image in a group of probable AD (n = 36) and control (n = 19) subjects with serial images. Using the method that gave most similar results to manual, three automated methods of calculating change within the hippocampal region were compared: fluid change calculated using (1) Jacobian change or (2) region propagation and (3) boundary shift. Rates were compared with manual measures. We found that segmentation of baseline hippocampus was most accurate using a template library combined with morphological operations (intensity thresholding plus one conditional dilation). This gave a voxel similarity of 0.69 (0.05) and 0.72 (0.06) in controls and probable AD subjects respectively compared with manual measures. Atrophy rates within these regions were most similar to the manual rates using the boundary shift integral (mean difference from manual rate 0.03% (1.29) in controls and 0.48% (2.44) in AD). A template library segmentation approach, together with morphological operations, provides a segmentation accurate enough to quantify relative change over time. The change over time can then be calculated automatically using boundary shift or fluid measures, with boundary shift giving most similar results to manual.

Introduction

Alzheimer’s disease (AD) is the most common cause of dementia worldwide. As the world’s population ages, the number of people with Alzheimer’s disease is set to increase from approximately 24 million today to 81 million by 2040 (Ferri et al., 2005). AD therefore presents an increasing socio-economic burden with each person affected representing both great personal loss and increased economic cost.

Currently, a definitive diagnosis of AD can only be made by histopathological examination of brain tissue, which is usually at postmortem. A positive diagnosis of AD can be made when the pathological hallmarks are seen: extracellular amyloid plaques and intracellular neurofibrillary tangles. One result of this pathology, progressive cerebral atrophy, can be visualized using structural imaging such as magnetic resonance (MR). The development and validation of noninvasive or minimally invasive markers of disease are important both to be able to aid diagnosis and assess the progression of the disease in the clinic and to assess a novel therapeutic agent in clinical trials where large numbers of scans require analysis.

Autopsy studies have shown the hippocampus to be affected by Alzheimer’s disease pathology early in the disease process, with approximately 20–50% loss of neurons by the time individuals are moderately affected (Bobinski et al., 1997, Braak and Braak, 1991). As a result imaging studies have focused on this region in order to test the efficacy of hippocampal atrophy as a predictor of AD (Korf et al., 2004, Scheltens et al., 2002). A large number of studies have shown that hippocampal volume is lower in AD subjects cross-sectionally (at one scanning time-point) although there is usually overlap with elderly controls owing to large inter-subject variability in volumes (Jack et al., 1992, Killiany et al., 2002, Krasuski et al., 1998, Laakso et al., 1995, Xu et al., 2000). Hippocampal volume change, usually expressed as a percentage loss per year, obtained from serial scanning (a number of time-points) means that every subject acts as his or her own control circumventing some of the problems associated with inter-subject variability in hippocampal size. Studies have shown that atrophy rates are significantly higher in probable AD than in age-matched controls (Wang et al., 2003); importantly, these rates of atrophy differentiate probable AD from controls better than absolute volumes (Barnes et al., 2004). Such measures of change may be more sensitive diagnostically and may also be useful in tracking disease progression or specific disease-related effects of treatment.

Manual delineation has been used for most MRI-based hippocampal volume studies whether longitudinal or cross-sectional (Chan et al., 2001, Fox et al., 1996, Jack et al., 1992, Killiany et al., 2002, Laakso et al., 1995) and this is currently considered to be the gold standard for hippocampal measurement. Manual voluming is both time consuming and requires trained operators making the use of this technique in large studies and clinical trials problematic. As a result, a number of semi-automated techniques have been developed to reduce the operator input. These differ in the levels of automation from those techniques that may require outlining of the hippocampus on a template image/atlas with region propagation/transformation to new images (Carmichael et al., 2005, Hammers et al., 2007, Heckemann et al., 2006) or propagation of a manually defined template onto the target image using manually defined landmarks on each scan (Csernansky et al., 2000, Haller et al., 1997). Other techniques use landmarks or manual intervention to allow the hippocampus to be segmented using either intensity and spatial information (Gosche et al., 2001), active appearance models (Duchesne et al., 2002), region growing (Chupin et al., 2007, Pitiot et al., 2004), or surface modeling (Ghanei et al., 1998). Some methods include a number of approaches; Shen et al. (2002) use a combination of manually derived landmarks, together with geometric and statistical priors to segment the hippocampus. Other methods considered to be entirely automated use both statistical and spatial information to segment a number of structures of which the hippocampus is one (Fischl et al., 2002).

Longitudinal methods are fewer in number and include those which require the manual outlining of the baseline hippocampus of a scan pair and measurement of shifts at the boundary of the hippocampus using the boundary shift integral (BSI) (Barnes et al., 2004) or non-linear registration and integration of the voxel compression map to estimate atrophy in the hippocampal region (Crum et al., 2001). Other methods result from the application of cross-sectional methods to longitudinal imaging including landmark-dependent template-based methods (Hsu et al., 2002, Wang et al., 2003) and four-dimensional template-based models which impose temporal smoothness constraints (Shen and Davatzikos, 2004).

One technique described the use of a manually delineated template which was linearly registered to the baseline images of a group of probable AD and controls with the template hippocampal region transformed using these transformation parameters (Barnes et al., 2007a). This template was chosen to be a subject from the study which had close to average hippocampal volumes for the wider group of probable AD and control subjects. This technique may be limited by the fact that the average volume may not be an optimal metric for template choice, and that this template may be suitable for some subjects, and less so for others. In addition, the boundary shift integral was used to quantify change over time in this hippocampal region, but it may be that change using a different automated technique for measuring change has greater agreement with manual measures or superior probable AD–control group separation. Further advances in template selection and calculation of rates of atrophy may be of benefit both for diagnostic purposes and for large studies where analysis of many hippocampi is required.

Our hypotheses were that a template library approach would prove to give more accurate hippocampal volumes compared with a single-person template and that simple morphological operations may further improve this accuracy. We also hypothesized that these segmentations combined with automated measures of change could be used to track atrophy progression over time. As a result our methodological objectives of this retrospective study were to assess whether (1) improvements could be made in the accuracy of template-based segmentations by use of a template that incorporates greater variability, (2) the resulting region could be made increasingly accurate using basic morphological operations, (3) the most accurate region could be used to quantify hippocampal losses over time using non-linear registration or linear registration combined with a boundary shift measure. The clinical objectives of this study were to assess whether automated measurement of hippocampal volume or rate are good diagnostic markers of probable AD.