Automated Labeling ofNeuroanatomical Structures inRoutine BrainCT

Although theadvanceinimagetechnology is remarkable, mostroutine brainimages arestill presented in series ofsections. We present a technique forautomatically assigning neuroanatomical labels toeachpixel inbrain CT sets basedonelectronic atlases. Incontrast toexisting segmentation procedures thatfocused onthin-cut MRI,ourtechnique could be applied onroutine CT images. Thetechnique employs anaffine registration formatching skulls togenerate template foreach slice. A deformable registration procedure isthenapplied to label theneuroanatomical structures. Theperformance ofour algorithm wasevaluated asareaoverlapping, measured atthe lenticular nuclei. Theaverage areaoverlapping ofimages without spaceoccupying lesions was72.3%. Withsmall space occupying lesions, this value is59.1%. Thetechnique isshownto beofacceptable accuracy forimages without spaceoccupying lesions. Itmightprovide basis forcomputer-assisted diagnosis andreport generation system. I.INTRODUCTION