Thirteen slices covering two patients who have been injected with Gadolinium and seven different scanning sessions were collected from a GE Advantage 1.5 Tesla MR imager coil. Each slice has a thickness of 5mm with no gap between consecutive slices.
Table 1: Comparison of Knowledge-Based Tumor Segmentation Versus
Hand Labeled Segmentation
After processing by our system, the final tumor segmentations were compared with supervised ``ground-truth'' tumor segmentations that were created by hand labeling. In all thirteen cases, the automatic tumor segmentation corresponded to the hand labeled image. There was some degree of error, false positives and negatives, as shown in Table 1, but since even ground truth is not absolute (no one can tell the exact tumor borders with non-intrusive methods), region correspondence is of primary concern, which this system handles quite well. Examples of automatic segmentation versus hand labeling are shown in Figure 5.
Figure 5: Comparison of Knowledge-Based Tumor Segmentation vs. Ground Truth.