Automatic Initialization of Contour for Level Set Algorithms Guided by Integration of Multiple Views to Segment Abdominal CT Scans Mahmoud Saleh Jawarneh 1 , Rajeswari Mandava 1 , Dhanesh Ramachandram 1 , Ibrahim Lutfi Shuaib 2 . 1 School of Computer Sciences, 2 Advanced Medical & Dental Institute Universiti Sains Malaysia, 1 11800 Minden, 2 13200 Kepala Batas, Penang, Malaysia {jawarneh, mandava, dhaneshr}@cs.usm.my, ibrahim@amdi.usm.edu.my Abstract—This paper presents a new automatic initialization procedure for a level-set based segmentation algorithm that works on all slices for a given CT dataset. Level set segmentation algorithms provide promising results, are robust to dataset variations and do not require prior training. As such, they can be reliably used for segmentation of major organs in abdominal CT scans. However, level set algorithms still require user intervention to plot the initial contour for each slice in a given dataset, which is a time consuming process. Therefore, we propose here, a technique of using multiple views to automatically initialize and propagate the contour through each slice in the CT dataset. The technique requires a user to only initialize a single point within the organ of interest in order to initiate the automated segmentation process. We report the segmentation results for liver and spleen organs within the abdominal region using three different datasets. We conclude that this technique can be used to reduce the processing time for any level set algorithm suitable to abdominal CT scans. We typically achieve time efficiency up to 203.03% for complete segmentation of three organs as compared to manually initializing the level set contour for each slice. Keywords-automatic initialization; mutliple views integration; level set active contour; abdominal CT images; medical knowledge. I. INTRODUCTION Segmentation of abdominal organs presents several challenges in CT images, showing high similarities in the gray levels among different organs and the surrounding soft tissues and inhomogeneity in shape and texture of organ tissues within and among different image slices. Additional difficulties are encountered when dealing with CT scans that have low contrast and blurred edges, due to partial volume effects resulting from spatial averaging, patient movement, beam hardening and reconstruction artifacts, as well as heartbeat and breathing [1]. The need for automatic segmentation of abdominal CT scans is ever growing. Currently, the segmentation is performed by experts who delineate organs borders on each slice in the volume dataset with the aid of semi-automatic tools. The result of segmentation in these processes depends on the skills of the operator in dealing with these tools and his/her knowledge of the target organ which suffer from his/her errors and biases, and also because these methods are tedious and time consuming [2]. There are several abdominal organ segmentation methods in CT images; including neural network learning techniques which most of these methods strongly depend on and require serious training set to build the shape and statistical or contextual constraints of organs to feed into neural network [3] [4]. Gray level based techniques such as thresholding, edge detection; mathematical morphology and region growing are required of the initial values, to start the segmentation and are based on intensity similarity. Over segmentation occurs when adjacent tissues have similar intensity to the target organ [2]. Rule-based recognition [3] on the other hand, is based on exploiting organ invariants and features such as size, location, edges and gray levels. Model-based techniques need training sets to be manually segmented and properly collected to produce the model correctly placed in dataset to give good results [2]. Atlas- based segmentation [5] [6] requires registration between built atlas and the target images. Active contour level set methods [7] [8] [9] give promising results, robust to dataset variations and not dependent on the training set. However they still require manual interaction from the user, to plot the initial contour inside specific organ. In addition, more computation is required for the level set algorithms to reach the desirable borders if the initial contour initialized farther from its final position. The focus of this paper is on the level set based active contour segmentation algorithms. And how to solve the problem of initializing the initial contour curve inside the target abdominal organ in each slice, near from the right organ’s border in CT abdominal dataset by using prior knowledge to help in automatic segmentation and to reduce the segmentation time. (a) (b) (c) (d) Figure 1. (a) Axial; (b) Coronal; (c) Sagittal; (d) All; views of abdominal dataset. Second International Conference on Computational Intelligence, Modelling and Simulation 978-0-7695-4262-1/10 $26.00 © 2010 IEEE DOI 10.1109/CIMSiM.2010.64 282 Second International Conference on Computational Intelligence, Modelling and Simulation 978-0-7695-4262-1/10 $26.00 © 2010 IEEE DOI 10.1109/CIMSiM.2010.64 315