Semi-automatic segmentation of MR and CT image sequences using a self-organizing map and texture descriptors

A system for semi-automatic texture segmentation of medical images, consisting of a self-organizing map (SOM) with a set of 16 texture descriptors as possible inputs, is presented. After a manually chosen calibration slice is used to train the SOM, the system can segment all the other slices from the same volume. It allows the selection of different combinations of descriptors providing adaptation to different image conditions. Also, a capability of labeling the SOM's outputs allows the merging or suppressing of segmented regions, providing extra flexibility for the segmentation. There is not previous knowledge about the tissues and the system detects as many regions as possible, because, at this moment, only fast segmentation is intended. The results show that although Principal Component Analysis (PCA) reveals second and higher order moments as the most discriminating features, these moments can be very sensitive to image variations and, therefore, segment excessive detail. And also that good classification results are obtained by using compact, local neighborhoods as small as 3 × 3 in medical images.