Model based validation scheme for organ segmentation in medical image data

Model based methods have proven to be among the most reliable and robust solutions for application of medical image processing in organ segmentation and reconstruction. They assume a repetitive geometric pattern for the organ of interest and therefore utilize probabilistic models to characterize organs shape or other attributes. A common strategy, adopted by most of the model based methods, is to use model attributes as prior information in the actual segmentation process. In this work, we propose a novel approach for accurate 3D organ segmentation and modeling in the CT scan volumes. Instead of direct use of the organs prior information in the segmentation process, here we utilize the knowledge of the organ to validate a large number of potential segmentation hypotheses that are generated by a generic segmentation process. For this, PCA based organ space is utilized to measure the fidelity of each segment to the organ. We detail applications of the proposed method for 3D segmentation of human kidney and liver in CT scan volumes. For evaluation purposes, the public database of MICCAI 2007 grand challenge workshop has been used. Implementation results show an average Dice similarity measure of 90% for segmentation of the kidney which shows better results than the 88.6% presented by other kidney segmentation methods. For the liver, the proposed algorithm achieves an average volume overlap error of 8.7% and an average surface distance of 1.51 mm while the best reported average values for these measures are 6.65% and 1.03 mm by an automatic algorithm on the same dataset.