Brain Studies and Validation

PD and T2 data sets were acquired axially on a GE 1.5 tesla MRI scanner, with repetition time TR = 2000 ms, and echo times of 35 and 70 ms, respectively. The slice thickness was 5 mm and the pixel size was 0.781 mm2. Each data set had 22 slices with 256 x 256 pixels per slice and was scaled linearly from the original 12-bit data to 8-bits. Figure 11 shows the intracranial boundary determined by the algorithm on selected slices of PD-weighted MR data where the boundary is accurate in all slices, except in slices 6 and 7, where the pituitary gland and basilar artery are included, and slice 5, where there is insufficient exclusion of the petrous temporal bone. The inclusion of pituitary gland and the petrous temporal bone generally do not affect subsequent analysis of the brain data.

Particularly in the high slices, the algorithm deals with the partial volume effects consistently.

Comparable results are obtained with our algorithm on more than 30 data sets from five scanners with fields of view varying from 200 to 260 mm. The algorithm also works on images acquired on a GE scanner with a SPRG sequence, with TR 39 msec and Te 8 msec, pixel size 1.0156 mm2, and slice thickness 2.5 mm. The computer processing time for each study for all the stages was less than 5 min on a SUN SPARC workstation — even the 120-slice 3D studies. In all cases, our algorithm detects the intracranial boundary without user interaction and without changing the parameters.

Tissue contours determined with a fully automated algorithm have to be validated with a study that compares them to contours traced manually by an expert. The similarity index described by Zijdenbos et al. [36], derived from the kappa statistic, can be used to compare an automated contour to one drawn manually. Each binary segmentation can be considered as a set A of pixels. The similarity between two segmentations A: and A2 is computed with a real number S e {0 ... 1} defined by l^i n a2| >iI + |a2

This similarity index is sensitive to both size and location since it depends on their intersection as well as the sum of their sizes. Two regions with equal size that overlap with half of their area have S = 1/2, whereas a region that completely covers a smaller one of half its size yields S = 2/3. In this manner, two regions where one fully encompasses the other are more similar than two partially overlapping regions. According to [36], good agreement is indicated by S>0.7, but the absolute value of S may be difficult to interpret. As

FIGURE 10 Refinement of the intracranial contour, (a) The contour defined by the perimeter of the initial brain mask, (b) The intracranial contour detected using the active contour model algorithm,

an example, the similarity index for the two images in Figs 12a and 12b is 0.942.

In a validation study that we conducted, three volumes were chosen and each volume was acquired using a different PD/T2-weighted echo sequence, and a different field of view size. For each volume, some axial slices were selected, such that the entire range of the image volume from "low" to "high" slices was covered. An expert radiologist traced the brain contour manually on each slice and the manual contour was compared with the automatically drawn contour using the similarity index. Table 1 shows the number of pixels included inside the manually drawn and automatically calculated brain contours as well as the similarity index.

In axial slices containing the eyes, the algorithm usually included the pituitary gland and basilar artery, and sometimes the internal carotid artery, whereas the radiologist excluded these structures (Fig. 11, slices, 6 and 7). Also, whereas the radiologist drew carefully around the petrous temporal bone, it was often included by the algorithm (Fig. 10b). Figure 12 illustrates one of these eye slices where the manual contour is in Fig. 12a and the automated contour is in Fig. 12b.

In the high slices, manual contours were comparable to the automated ones except in the extreme case of the top slice of the 5 mm thick datasets where the partial volume effect was noticeable. The sagittal sinus was usually included by the algorithm, whereas it was always excluded by the radiologist (Fig. 11, slices 18, 20, and 22).

Overall, this algorithm provided a similarity index always above 0.925, it was maximal 0.99 on middle slices, and dropped to 0.95 on the highest slices. These results compare favorably with those reported by others [19,1] as the brain volumes are within 4% in most cases.

0 0

Post a comment