Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems
Dequan E. Medina - Fort Valley State University
Co-Author(s): Dr. Chunhua Dong, Fort Valley State University, GA
Accurate segmentation of organs in CT images is very important in computer-aided diagnosis. The learning-based organ segmentation can be viewed as a classification problem that aims to differentiate target objects from other tissues. In this study, we propose a novel segmentation method which utilizes the similarities of intensity distribution, shape and location between nearby CT slices as a priori knowledge to guide the segmentation of a volume of image sequence, where the classification is performed using Support Vector Machines (SVM).
As the first step of our strategy, we manually draw the region of interest from the first CT slice image and select training data for SVM training. To enhance the contrast of the CT image, we extract statistical intensity information from the training data and model the object intensity distributions using a Gaussian function. After the pre-processing of applying the Gaussian model to the image pixels, we obtain the organ segmentation result of the first slice using SVM.
Due to the similarities of organ shape and location between adjacent slices, the segmentation result of the first CT slice in the volume can be used as a priori knowledge of the organ in the adjacent slice. According to this a priori knowledge, the background and object training data of the adjacent slice can be automatically obtained by enlarging and shrinking the identified boundary in the first slice. Following a similar pre-processing method, we use this training data for the SVM learning process and finally classify the organ from other tissues in this adjacent slice. This process can be repeated over the image sequence using the new segmentation results as a priori knowledge for the next slice. Experiments have been conducted of extracting the liver from the clinic images. This method yielded accurate segmentation results, comparable to ground truths. Our future research would be to explore the application of other classification methods for CT images.
Funder Acknowledgement(s): This study was supported by a grant from the Department of Defense awarded to Xiangyan Zeng PhD, Professor of Computer Science at Fort Valley State University.
Faculty Advisor: Xiangyan Zeng, zengx@fvsu.edu
Role: The part of the research that I conducted included the implementation and use of SVM on the CT Images, with the help of my adviser. Before using SVM, I wrote and utilized the Gaussian function, in order to receive better results from SVM.