Development Of Automatic Liver Segmentation Method For Three- Dimensional Computed Tomography Dataset

• Main Author: Chew, Chin Boon
• Format: Monograph
• Language: English
• Published: Universiti Sains Malaysia 2018
• Subjects: T Technology; TK Electrical Engineering. Electronics. Nuclear Engineering
• Online Access: http://eprints.usm.my/53383/1/Development%20Of%20Automatic%20Liver%20Segmentation%20Method%20For%20Three-%20Dimensional%20Computed%20Tomography%20Dataset_Chew%20Chin%20Boon_E3_2018.pdf; http://eprints.usm.my/53383/

Computer tomography (CT) is usually used as the medical imaging modality for liver. Liver segmentation is important as it is preliminary for liver diagnosis. Manual segmentation can provide good result based on the skill of radiologist but the process is tedious and time-consuming due to large number of slides produced by the CT scanner. Many researchers try to develop and proposed various liver segmentation methods which can be classified into automatic and semi-automatic segmentation. Both methods are able to speed up the segmentation time. The low contrast of liver boundary with neighbouring organs, high shape variability of liver, presence of noise in image and presence of various liver pathologies make liver segmentation very challenging. Despite that, automatic segmentation is still the more desirable method to be use due to its efficiency and convenience. Therefore, an automated liver segmentation algorithm is proposed and developed in this project to improve the accuracy and time required for liver segmentation. The proposed algorithm can be divided into two parts. The first part is to build the probabilistic atlas. Probabilistic atlas provides the probability for a voxel to be a liver and act as a guide for segmentation. Both the location based and intensity based probabilistic atlas are built from the 20 datasets obtained from SLIVER07. The atlases act like a guide for segmentation. The second part is to use the probabilisticatlas built to segment the liver. The segmented liver will be refined by the probabilistic atlases itself and then further refine by morphological closing and 3D median filter. The proposed algorithm is then tested by the 19 datasets that are used to train the atlases as the ground truth datasets are required for evaluation. The evaluation on the performance is based on volumetric overlap error (VOE), relative volume difference (RVD) and dice similarity coefficient (DSC). The proposed algorithm provided mean VOE of 26.50%, mean RVD of 15.09% and mean DSC of 0.8421. The time required for segmentation is 366s. The segmentation results from the algorithm developed are competitive. However, improvements still can be made.