Abnormal slices are much more difficult to
label because (a) tissue shapes, such
as white matter and CSF, have been deformed to various
extents by the intrusion of tumor masses; (b) abnormal
tissues like tumor, edema and necrosis are very close in
intensity features to their neighbors, CSF and gray
matter.
A. Labeling Gray Matter:
For most abnormal slices, the labeled tissues include
skull tissues and white matter.
The remaining classes belong to gray matter, CSF, and
abnormal tissues. Since the original slice has been
over-segmented, several classes may correspond to one
tissue. Figure 3(a) shows an image formed by
merging the unlabeled classes of a slice, which are
regions of interest (ROI). We have already identified
and labeled the clusters corresponding to air, skull
tissues, and white matter, so we know that the lowest
class of the remaining unknown classes in the T2 spectrum
belongs to gray matter, as shown in
Figure 3(b). Since other tissues might
also be clustered into this class, which occupies
the space around CSF, only pixels with strong certainty
belonging to gray matter will be chosen from the cluster
to be gray matter. A brain region is formed by merging
white matter tissue and the ROI classes.
The pixels of strong certainty are selected by
subtracting the smaller brain region from the gray matter
class (Figure 3(c).)
These pixels are used to initialize the FCM algorithm,
which clusters three feature images
masked by the ROI classes into two classes: gray matter
and CSF plus abnormal tissues. Figure 3(d)
shows the clustered results. Figures 4(a)
and (b) show two gray matter classes of two overlapping
slices before and after the injection of gadolinium, a
tumor-enhancing agent. Figure 4(a)
consists of gray matter and another tissue, a blob at the
lower left horn of CSF, which is more sharply defined than in
Figure 4(b). The number of
pixels in Figures 4(a) and (b) are 8085 and
7300 respectively. After
reclustering as above, these two classes are almost
identical, as shown in Figures 4 (a') and
(b'), with number of pixels being 8277 and 8397
respectively. Since the slices do
cover only nearly the exact same anatomical region,
some small difference is to
be expected.
B. Labeling CSF and Abnormal Tissues: Thirty abnormal slices have been tested. In all segmentations gray matter, white matter and CSF are correctly labeled. Identifiably correct labeling of all tissues has been achieved in eleven slices. The other 19 slices appear to be segmented correctly, but have not been compared with patient history and dissected by the attending physicians. In many cases it is difficult, even for expert radiologists, to label abnormal tissues without reference to patient records. The method currently used to select abnormal pixels is most effective for slices with significant abnormal tissues.
Figure 3: Reclustering to label gray matter.
Figure 4: Gray matter before and after reclustering