Distinctive patterns of accelerated cerebral atrophy are a feature of a large number of neurodegenerative and dementing disorders such as Alzheimer's disease, frontotemporal dementia, Parkinson's disease and others. Therefore, analysis of these patterns can be used for diagnostic decision support. Cerebral atrophy can be measured either through measurements of grey and white matter volumes, or measurements of the volume of the cerebro-spinal fluid (CSF). Since the interior volume of the cranium is approximately fixed throughout adulthood, increases in CSF volume provide an accurate marker for decreases in grey and white matter volume. However, the CSF has markedly different physical properties to the other tissues present in the cranium, and so it is possible to design MR pulse sequences that exhibit high grey-level separation between CSF and the other tissues present. This leads to more accurate segmentation, and thus more accurate volume measurements, than would be possible if grey and white matter volume were measured.

The DODECANTS tool implements an updated version of the functionality described in [8]. The aim of the algorithm is to map the distribution of CSF volumes in the cranium at a coarse level, and to perform nearest-neighbour classification of the distributions in previously unseen dementia patients through comparison with the distributions in a set of patients for whom reliable diagnoses are available. These classifications can then be used for diagnostic decision support. Since the interior volume of the cranium is approximately fixed throughout adulthood, increases in CSF volume provide an accurate marker for decreases in grey and white matter volume. However, since the grey-level separation between CSF and the other tissues present is higher than the separation between grey/white matter and bone, fat and air, for the MR imaging sequence used here, the signal-to-noise ratio for the segmentation process is higher and thus the segmentation is more accurate.

The DODECANTS tool combines functionality from several other TINA tools. The analysis procedure consists of four stages: registration, segmentation, CSF volume measurement and normalisation, and nearest-neighbour classification. First, all images under analysis are registered to a consistent coordinate system using rigid Mutual Information (MI) based registration [1,2], as described in TINA Memo nos. 2001-013, 2003-002 and 2004-001. Segmentation of the CSF is achieved using the EM-based algorithm described in TINA Memo no. 2004-009 [7] (see also [6], [4], [10], and [5] for details of the development and testing of this algorithm). The CSF volume maps are then multiplied with a set of binary masks, drawn by hand in the standard coordinate system. The masking has two purposes: to delete non-CSF fluid spaces (e.g. eyes, sinuses) and to enforce a consistent inferior boundary to the measurement space, defined by drawing a line in the mid-sagittal section parallel to the horizontal axis that passes through the junction of the calvarium and the tentorium cerebelli. The anterior, posterior, lateral, and superior boundaries of the CSF space are automatically identified by locating the extremes of the CSF.

The CSF space is then divided into twelve equi-sized rectangular volumes, defined by planes which divide the space into anterior, central and posterior thirds, lateral halves, and superior and inferior halves, and the CSF volume in each of these regions is measured by counting the number of pure CSF voxels, and the number of partial volume voxels containing more than 50% CSF by volume, and multiplying with the voxel dimensions. The purpose of the arbitrary division of the space is to allow analysis of patterns of cerebral atrophy independent of any prior hypothesis regarding spatial distribution. The volume measurements are then normalised for variation in head size and for normal age-related atrophy, using the procedures described in TINA Memo no. 2004-002 [3]. Head size normalisation is accomplished by dividing all volume measurements by the total intracranial volume (TIV), computed from the volume of the bounding box on the CSF space as described in TINA Memo no. 2004-002, which is in turn computed by finding the product of the maximal extents of the CSF in the anterior-posterior, lateral, and superior-inferior directions. Normalisation for age-related atrophy is accomplished using a Weibull cumulative distribution functional fit to volume measurements from 70 normal volunteers, as described in TINA Memo no. 2004-002.

The twelve normalised CSF volumes are then used to calculate five reduced variables,

where F is the sum of the anterior four volumes, M is the sum of the central four volumes, B is the sum of the posterior four volumes, P is the sum of the left six volumes, S is the sum of the right six volumes, U is the sum of the superior six volumes, and L is the sum of the inferior six volumes. Left and right in this case are the left and right sides of the image, not necessarily of the patient. These variables are relative, and so errors in certain stages of the analysis, such as misregistration, will be removed to first order by this procedure. Finally, a k-nearest-neighbour classifier is run on the reduced variables, using a leave-one-out approach to classify each data set on the basis of the others. These diagnoses can be used to estimate the sensitivity and specificity of the technique.

It should be noted that the DODECANTS tool is still under active development. Therefore, the interface has been optimised for maximum functionality, rather than simplicity, and so some of the procedures described below are complex. In addition, checks have not been implemented for all memory-related failure modes (e.g. missing input files), and so the tool may crash if misused.