Automatic Segmentation of non-contrast-enhanced Meningioma in MR Images Using Combined Fuzzy c-Means, Region Growing and Knowledge-Based Techniques

• Main Authors: Yi-Min Liu, 劉怡旻
• Other Authors: Jau-Min Wong
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/30660727251176361887

碩士 === 國立臺灣大學 === 醫學工程學研究所 === 98 === In recent years, magnetic resonance imaging (MRI) has became an important modality for brain tissue diagnosis, due to its high resolution, less radiation injury and an excellent resolution for soft tissue imaging. The tissue characteristics can be described by different pixel intensity in the feature space of T1, T2 image via MRI system. This is especially useful for any attempt to segment brain tissues, normal or abnormal in clinical practice. Tumor segmentation is a key work for brain disease diagnosis. However, manual brain tumor segmentation from magnetic resonance images is a cumbersome and time-consuming task for physicians. For this reason, an automated brain tumor segmentation method is desirable. This system used two non-contrast-enhanced MR images, T1 and T2 image to segment brain tumor automatically. In the beginning, we performed a multi-spectral histogram for pixels clustering by FCM, and merged the neighborhood pixels of seed by region growing. By knowledge-based information, our system selected tumorous clusters and merged them into one tumorous image. Finally, we optimize the tumorous image with morphology technique. Based on above algorithm, this system used only two non-contrast-enhanced T1 and T2 MR images to label tumor and measure the area of tumor automatically. To evaluate the segmentation results, they were then compared to “ground-truth” (GT) on a pixel level. The performance in automatic brain tumor segmentation of our system, the total PM varies from 15.5% to 99.91% with a mean and standard deviation of 86.98% and 17.42% respectively, and the total CR varies from 0.16 to 0.95 with a mean and standard deviation of 0.782 and 0.169 respectively. While in the semi-supervised brain tumor segmentation of our system, the total PM varies from 15.5% to 99.91% with a mean and standard deviation of 87.85% and 15.93% respectively, and the total CR varies from0.16 to 0.95 with a mean and standard deviation of 0.793 and 0.155 respectively. This statement represents that tumorous pixels were not only highly match between GT and our system, but has a fair level of correspondence between GT and our system. Therefore, this system has a great potential of becoming a clinical MR images analysis tool for helping experts to obtain tumor location and volumetric estimation or eventually to therapeutic planning in the future.