| Summary: | 碩士 === 國立中正大學 === 資訊工程所 === 95 === Assessment of newborn cerebellar volume is an important process for detecting and diagnosing cranial nerve diseases; therefore, a proper treatment would be given if early detection is found. In this paper, an implicit model-based semi-automatic level set fast marching system is proposed to segment the 3-D cerebellum structure in the cranial US image. In this system, only the segmentation seeds of the first image have to be marked manually and the seeds of the other images could be found by the segmentation result of the previous images automatically; after that, a 3-D cerebellar volume could be reconstructed from those 2-D segmented images. Due to low contrast and speckle noise in the US images, there are four pre-processing procedures exploited. The anisotropic diffusion is adopted to smooth the image, the contrast enhancement is used to distinguish among the cerebellum and other structures, the gradient computing is utilized to determine the object contours and discriminate from homogeneous regions, and the sigmoid filtering is employed to enhance the object boundary. Moreover, the 3-D morphological closing operation with a 3×3×3 ball structuring element is applied for reducing discontinuity of any two continuous images. Finally, there are total 75 cases examined and a commercial 4-D view program is adopted to calculate the cerebellar volume for comparing with our results. The proportion of the experimental results with the match rates smaller than 5% is over 90% and more than one third of experimental results with the match rates smaller than 1% are obtained. However, there are still 7 worse cases with the match rate larger than 5% due to the grey levels between the cerebellum and surrounding structures are too similar. To solve this problem, a proper ROI containing only the cerebellum could be selected to improve the segmentation result effectively. The experimental results show a high accuracy with our developed system.
|
|---|
