Development of an inductive self-organizing network for the real-time segmentation of diagnostic images

A self-organizing neural network has been developed far the application to the segmentation of magnetic resonance images, in particular to the distinction between white and gray matter, useful: to the computer guided neurosurgery. By directly comparing the images of a patient with an on-line anatomical atlas, the guided surgery system is able to identify the exact position of the sensor placed and moved into the brain. The developed neural network (ITSOM) derives in real time the correct value of a pixel coming from a data stream starting from a narrow reference set through the comparison between the cyclic attractors generated by the set itself and by the input data In this way it has been possible first to classify a set of gray levels from MR T2 scans, then to give each class the white/gray matter attribute by inductively comparing the features of such classes with the features of known pixels. The system is able to segment a 256x256 pixel image in no time an Pentium 133, adapting itself to the strict time needs of the operating room.