A set of fuzzy rules whose antecedent fuzzy sets are adapted to each each image are shown to be effective in reducing the time to segment magnetic resonance images of the human brain into tissues of interest. The segmentation produced by the fuzzy rules serves as an initialization to a semi-supervised clustering algorithm which produces the final segmentation. The developed hybrid segmentation system is approximately 5 times faster than fuzzy c-means clustering. It has been tested on 105, 5mm thick, magnetic resonance image slices of the human brain using T1, T2, and proton density weighted images as feature images (i.e. each voxel has 3 features). The images come from 15 different subjects and span a range from the ventricles (roughly the middle of the brain in the axial plane) to the top of the brain.
The hybrid segmentations are insignificantly different than those obtained with fuzzy c-means clustering when compared with a pseudo ground truth created from a supervised k nearest neighbor segmentation.
The overall performance of the segmentation approach demands further refinement using some kind of knowledge. An example of a slice with significant extra-cranial tissue that is mis-classified in this approach as white matter is shown in Figure 10. Since the tissue is clearly outside the skull, simple knowledge about removing all tissue spatially outside the skull would prevent this tissue from being considered during processing. Such knowledge guided approaches have been applied in other work [4,2].
The approach of using fuzzy rules whose antecedents fuzzy sets are created from intensity histograms can be applied to other domains of images taken of the same region over time as long as the shape of the histograms remains approximately constant. Such rules provide a fast initial segmentation that can be further refined via other image processing techniques or with the use of heuristics in conjunction with image processing algorithms.
| Normal | Abnormal | |
| Test | 1.4 | 0.82 |
| Train | 0.48 | 0.29 |
