Over the past decade and half a large literature has developed on the topic of computer aided segmentation and quantitation of brain MRI. Much of this literature has focused on algorithms designed to measure global brain and CSF volume that can be accomplished with high reproducibility(24
). It is generally acknowledged, however, that the accuracy and reproducibility of automated image segmentation is much worse for small, spatially non-contiguous tissues like leukoaraiosis than for larger structures like the whole brain(26
). Like leukoaraiosis, MS plaques are most often located in the white matter, typically have elevated T2 signal with respect to adjacent normal brain tissue, and assume an anatomic configuration either of spatially distributed discrete foci or confluent areas of elevated signal in the periventricular white matter. In many respects therefore, quantification of leukoaraiosis volume is more analogous to MS quantitation than to whole brain quantitation; and the literature on automated quantitation of MS plaque volume is far larger than the literature on automated quantitation of leukoaraiosis volume(33
Multi-spectral segmentation algorithms are a commonly used approach to the problem of segmentation, particularly of MS plaques(34
). Two or more spatially registered MR image volumes with different contrast properties are used to define a feature-space. The images are segmented into different tissue classes based on the principle that specific tissues form clusters in feature-space. An advantage of the FLAIR-histoseg algorithm is that tissue classification is accomplished with only a single image volume. This obviates the need for registration of multiple image datasets. While registration algorithms generally work well, it does require an additional step that is unnecessary with FLAIR-histoseg or other single-band algorithms. Also, situations do arise in which registration functions suboptimally. For example, in order to achieve adequate spatial resolution in the slice select direction, and minimize partial volume averaging, an interleaved acquisition with thin sections (≤3 mm) is recommended for MS plaque (and leukoaraiosis) quantitation(37
). A situation which we frequently encounter in our elderly and demented subjects is that one of the "packs" of the interleaved acquisition is slightly out of registration with the others due to patient head motion during the acquisition. Often this is inplane axial head rotation. If the inplane axial rotation occurs during the acquisition of non-centric phase encoding views, the mis-registered slices may not be excessively blurry, and are therefore "usable", just mis-aligned with respect to the images in the other interleaves. In this situation, rigid body multi-image registration required for many multi-spectral classifiers would be invalid. However, by obviating the need for registering multiple image volumes, the FLAIR-histoseg algorithm can estimate leukoaraiosis lesion volume even in the presence of some intra-image rotational motion.
A second advantageous feature of the FLAIR-histoseg algorithm is that the actual classification of pixels operates unsupervised, in the sense that no operator input is necessary to determine tissue intensity values used by the algorithm to segment the image. (Operator intervention is required in the pre-processing steps, so the algorithm as a whole is best regarded as semi-automated). The FLAIR-histoseg algorithm was "trained" in advance using phantom image datasets in which correct tissue classification was known. Supervised algorithms on the other hand require a trained operator to manually identify training sets of the major tissue classes of interest, for example CSF, brain, and leukoaraiosis lesion for each new set of images. The final result of a supervised classifier is highly dependent on operator defined tissue classification input values which are unique to each dataset(26
). Small differences in operator judgment about the training dataset(s) may produce wide variation in results (32
Surface erosion was employed in preprocessing, because we noticed that in FLAIR images the limbic cortex often had increased signal relative to white matter and neocortex. As a result, without surface erosion, some limbic cortex was mis-classified as lesion. In a recent publication, Hirai et al (22
) confirmed this observation of increased signal intensity of limbic cortex relative to neocortex. Even with the surface erosion step, some manual editing of pixels in the limbic cortex that are incorrectly classified as lesion is usually required. Manual editing is also necessary for flow-related artifacts, pixels in the posterior limb of the internal capsule which often have high signal intensity, and pixels in central grey nuclei with dense physiologic mineral deposition which have low signal intensity (23
). These manual editing steps are the main source of test re-test variability in the method. Manual editing is also used to segment the brain from the skull and scalp. This step could be automated to some extent using erode/dilate procedures. However, even with putatively fully automatic algorithms, some manual "clean-up" is usually required in order to obtain pristine brain segmentation.
The method we developed for creating synthetic image phantoms of leukoaraiosis allowed us to first develop the segmentation algorithm and then later, using a different set of image phantoms, test the accuracy of the algorithm with respect to a known gold standard measure. The leukoaraiosis lesions introduced into the synthetic images had signal intensity properties and spatial distribution properties that were fairly representative of that found in real life.
The FLAIR histoseg algorithm operates on the following principal. Pixels which represent normal brain can be accurately identified by statistic evaluation of the central portion of the histogram of FLAIR images, and pixels whose intensity exceeds that of normal brain represent lesion. We chose to use the mode value of the histogram to identify the central value of normal brain pixels. Unlike other measures of central tendency, the mode does not fluctuate with variations in the magnitude of the values in the tails of the histogram, which by definition vary across slices and across patients. Smoothing the histogram prior to determining its mode reduced variability associated with using the mode as the measure of central tendency.
The mean measurement error of FLAIR-histoseg algorithm with respect to the gold standard leukoaraiosis volume embedded in a number of different phantom images was 6.6%. This compares favorably with the accuracy values published for various other classifiers which range from 2.2% to 20%, using either synthetic images or physical phantoms as the measure of absolute truth (38
). Because designing anatomically realistic phantoms of MS or leukoaraiosis is difficult, most segmentation algorithm validation studies have addressed test re-test reproducibility rather than measurement accuracy compared to an absolute volumetric gold standard. With a mean test re-test co-efficient of variation of 1.4%, the FLAIR-histoseg technique compares favorably with reproducibility assessments of other algorithms for MS and leukoaraiosis quantitation. Published reproducibility measurements expressed as coefficients of variation range from 0.9% to 39% (35
In summary, FLAIR-histoseg is a viable method by which the volumetric burden of leukoaraiosis can be measured in elderly persons. The measured levels of accuracy and reproducibility can be achieved by simply thresholding a single image, due to the unique contrast properties of the FLAIR pulse sequence which isolates the pathology of interest to a specific portion of the intensity histogram. This avoids the need for multi-image registration. The unique aspect of the segmentation algorithm itself centers on automated statistically based computation of histogram thresholds for tissue classification, which are determined individually for each slice and thus accommodate slice to slice variation in the intensity histogram distribution.