Semiautomated and Automated Methods

Semiautomated and automated segmentation programs offer more rapid assessment of areas and volumes. Such programs have been used to calculate spinal cord cross-sectional area (Losseff et al., 1996b), lateral ventricle volume (Matthews et al., 1996; Lycklama et al., 1998; Brex et al., 2000; Kalkers et al., 2001), CSF volume (Wolinsky et al., 2000; Dastidar et al. 1999), and whole or partial brain volumes (DeCarli et al., 1992; Losseff et al., 1996a; Bedell and Narayana, 1996; Hohol et al., 1997; Fisher et al., 1997; Goldszal et al., 1998; Alfano et al., 1998; Phillips et al., 1998; Rovaris et al., 2000; Collins et al., 2001). There are many different site-specific segmentation algorithms and programs currently in use for MS applications.

One example of a semiautomated technique is a threshold-based algorithm that was developed to measure spinal cord cross-sectional area (Losseff et al., 1996b). An operator selects regions of interest to determine an intensity threshold midway between CSF and spinal cord tissue. To minimize axial repositioning errors, the cord is segmented in five adjacent image slices, and the area is determined as the mean cross-sectional area. Scan-rescan tests resulted in mean CVs ranging from 0.79% to 1.6%. A similar threshold-based approach has been applied for semi-automated segmentation of the lateral ventricles in T1-weighted images (Brex et al., 2000; Luks et al., 2000). First, the mean intensity of brain parenchyma is determined from automated segmentation or operator-selected regions of interest, and then the threshold is determined as 60% of the brain intensity. The intrarater CV for this technique is 0.13% (Brex et al., 2000) and the intraclass correlation coefficient is 0.99 (Luks et al., 2000).

Various methods have also been applied to measure brain parenchymal volume. Brain segmentation typically consists of two steps: (1) separation of tissue from CSF and background, usually by intensity thresholding; and (2) separation of the brain tissue from other cranial structures, usually by connectivity and morphological operations, manual interaction, and/or knowledge-based anatomical operations. An example of a semiautomated algorithm allows the user to interactively choose low and high thresholds that cover the intensity range of brain parenchyma, and then select a seed point within the parenchyma on a slice-by-slice basis (Rovaris et al., 2000). A region is automatically grown around the seed point that includes all connected pixels within the given range of intensities. Boundaries are drawn manually when necessary to prevent the region from growing outside the brain and into other structures. The intraob-server CV for this technique calculated from resegmentation of 10 images is 1.9% for whole brain volume.

Another approach is to perform the same steps automatically (Fig. 2): (1) Automatically determine the optimal threshold for separation of parenchyma and CSF based on histogram analysis or multispectral classification; (2) apply morphological operations to erode small connections between the brain and other cranial structures; and (3) use connectivity principles to identify the largest connected component within the image. Variations on these steps have been implemented by several groups to generate an initial segmentation of the brain (Losseff et al., 1996a; Bedell et al., 1996; Hohol et al., 1997; Fisher et al., 1997; Goldszal et al., 1998; Collins et al., 2001). In general, however, after the third step there are still some nonbrain structures included in the segmentation, and additional processing is required. In one variation on this approach, the segmentation is restricted to a central 20-mm thick slab of tissue selected by the radiologist (Losseff et al., 1996a). Manual editing is performed after automated segmentation, if necessary. The CV is 0.56%, as determined by a scan-rescan test. The same basic approach has also been applied to whole brain segmentation (Bedell and Narayana, 1996; Fisher et al., 1997; Goldszal et al., 1998; Collins et al., 2001; Udupa et al., 1997; Kikinis

Figure 2 Generic automated brain segmentation algorithm: (A) slice from the original PD/T2 dual echo image (early echo minus late echo); (B) optimal thresholding to separate tissue from background and CSF; (C) morphological erosion with a 5 x 5 x 5 diamond-shaped kernel to disconnect connected structures; (D) identification of the largest connected component, the brain parenchyma; (E) morphological dilation with a 5 x 5 x 5 diamond-shaped kernel to recover the original shape; (F) boundaries of final segmentation superimposed on the original image.

Figure 2 Generic automated brain segmentation algorithm: (A) slice from the original PD/T2 dual echo image (early echo minus late echo); (B) optimal thresholding to separate tissue from background and CSF; (C) morphological erosion with a 5 x 5 x 5 diamond-shaped kernel to disconnect connected structures; (D) identification of the largest connected component, the brain parenchyma; (E) morphological dilation with a 5 x 5 x 5 diamond-shaped kernel to recover the original shape; (F) boundaries of final segmentation superimposed on the original image.

et al., 1999) and implemented as semiautomated programs (with only minor editing requirements) or fUlly automated programs. Measurement variability with these techniques is consistently below 1%.

An important distinction between volumetric methods of atrophy measurements is whether the structure size is reported as the actual volume (e.g., in milliliters) or as a normalized volume. Normalized measures of whole brain atrophy are calculated as the brain parenchymal volume divided by an estimate of the intracranial volume to correct for head size. Normalized brain volume calculated in this way is referred to as the brain-to-intracranial-cavity volume ratio (Hohol et al., 1997), percent brain parenchyma volume (PBV) (Phillips et al., 1998), brain parenchymal fraction (BPF) (Rudick et al., 1999), brain-to-intracranial-cavity ratio (Collins et. al. 2001), or parenchymal fraction (Kalkers et al., 2001). CVs for normalized brain volume range from 0.2% to 2%, depending on the level of automation in the segmentation (Kalkers et al., 2001; Rudick et al., 1999). Head-size normalization is particularly important in cross-sectional studies in which normal biological variation in head size can easily obscure subtle disease-related volume differences. In normal healthy controls, normalized brain volume is fairly consistent between the ages of 20 and 55 years (Pfefferbaum et al., 1994). Therefore, normalized brain volumes also provide a means to estimate the total amount of atrophy that has occurred up to the time of the scan compared to an age-matched healthy control group. Normalization is also important in placebo-controlled longitudinal trials, where it is necessary to establish that two groups of patients are comparable at baseline. Using absolute brain volumes, it is not possible to ensure that the placebo group and the treated groups do not have different amounts of atrophy at the start of a trial.

Another type of whole brain atrophy estimation calculates brain atrophy directly from images acquired serially over time (Fox et al., 2000; Smith et al., 2001). These techniques require accurate co-registration of the images to determine changes in the brain-CSF boundaries over time. The mean error in atrophy measurement using this technique is approximately 0.2% (Fox et al., 2000).

It is important to note that atrophy measurements based on MRIs may be affected by factors unrelated to MS disease processes. Technical factors include patient positioning, scanner hardware and software upgrades, partial volume effects, motion artifacts, dental artifacts, voxel size calibration, intensity inhomogeneities, protocol or sequence variations, and the like. The measurement strategy may circumvent some of these problems. For example, one- and two-dimensional measures require precise patient repositioning. Volumetric, or three-dimensional techniques that use image registration to the baseline scan, on the other hand, reduce the effects of patient repositioning on atrophy measurements. Scanner upgrades and voxel size drift can be partially corrected using phantoms and consistent calibration procedures during longitudinal studies, or by using normalized measures of atrophy. Decreasing slice thickness or accounting for partial volume effects in volume calculations may minimize partial volume effects. Atrophy measures are affected by aging, with decreasing brain volumes observed in populations of healthy individuals over age 50. Atrophy measures may also be affected by alcohol ingestion, dehydration, concurrent CNS disease (e.g., cerebral vascular disease), and medications (e.g., steroids, diuretics). It is important that groups be age-matched, and that exclusion criteria include confounding medication and coexisting brain disease.

Was this article helpful?

0 0

Post a comment