This paper describes a new method for extracting and segmenting intracranial structure from serial images of cerebral computerized tomography automatically. A region growing- and morphology-based approach was first developed to extract intracranial structures from the serial images of cerebral computerized tomography, and focusing on the problems of parameter initialization of the expectation maximization (EM) algorithm, an improved EM algorithm based on parameter- limited GMM was presented to segment the intracranial structures successfully. Experimental results of the algorithm showed that this method was effective for all cerebral computerized tomography images from bottom to top of the cerebrum.