ENHANCED AUTOMATIC COLON SEGMENTATION FOR BETTER CANCER DIAGNOSIS Marwa Ismail 1 , Aly A Farag 1 , Robert Falk 2 , and Gerald W Dryden 3 1 University of Louisville, Louisville, USA 2 Medical Imaging, Jewish Hospital, Abraham Flexner Way, Louisville, USA 3 Division of Gastroenterology/ Hepatology, University of Louisville, USA ABSTRACT Colon segmentation is the first stage towards polyp detection, the main cause of colon cancer. Due to the immense importance of colon cancer diagnosis which is the second leading cause of death in the world, the segmentation phase must guarantee that no polyps are missed, especially the flat ones that are usually hard to detect. This work validates the 3D automated colon segmentation approach using the convex contour model previously proposed in literature. It also adds improvements to its pre-processing stage in order to better capture the colon walls and to enhance the results of the subsequent phases of the segmentation process. Experiments were conducted on 27 colon data sets that include 30 polyps. Moreover, 30 synthesized polyps with various shapes and sizes were placed at challenging areas of the colon’s complex structure. Experiments conducted show a significant improvement in the construction of colon walls and the rate of polyp detection over that provided by the original technique. Index Terms— air-filled colon, fluid-filled colon, non- colonic attachments- polyps-convex snake model 1. INTRODUCTION Virtual Colonoscopy (VC) has become one of the most reliable techniques for polyp detection, which is the primary cause of colon cancer. It has a growing interest among physicians with colon cancer being the second leading cause of death in the world. The first stage of a VC procedure is colon segmentation from abdominal CT scans that will significantly affect subsequent VC stages if it is not accurately employed. Poor colon segmentation would dramatically affect the rate of polyp detection, especially for flat ones that grow directly to the colon wall. Colon segmentation itself is a highly challenging problem due to many reasons including: 1) the presence of other structures in CT scans with the same intensity as of air-filled colon segments, such as lungs, and small bowels. 2) The tortuosity of the colon (caused by haustral folds). 3) The partial volume (PV) effect caused by residing of the contrast agent, the colon is injected by to insufflate it, in the lower concave parts of colon segments. An air-fluid boundary (AFB) region will then develop that will disconnect the colon into two parts, fluid-filled (similar to bone in intensity), and air-filled. 4) Poorly distended colon sets are collapsed at some areas, which will make false interpretation of colon segments as small bowels [1]. Colon segmentation has been widely addressed in literature. In [2], pre-processing algorithms, followed by 3D region growing were employed. In [3], a multistage approach was developed with two levels of classification based on vector quantization and region growing. A post-processing method was proposed in [4] to repair the gaps in the segmented colon. In this paper, we improve our work proposed in [5] for automatic colon segmentation that is based on the convex formulation of the active contour model. Pre-processing stage is of great importance, as it deals with the partial volume effect that might be the cause of missing significant colon parts if not well handled. We improve the pre-processing stage of the previously proposed algorithm and validate it with an extended number of real and synthetic polyps, which was not provided at our previous work. Results show improvement in the rate of polyp detection from the original algorithm, especially for flat ones that were missed before. The rate of poly detection is up to   %. 2. METHODS Our work in [1] had 3 stages: 1) digital cleansing for detection of AFB and merging air and fluid colon parts. 2) air-filled parts segmentation based on [6]. 3) post processing to remove structures with the same intensity of that of the colon. In this paper, we enhance phase 1 of this framework for the sake of detecting all colon walls and better detection of polyps. Figure 1 (a) shows a CT slice with PV that disconnects the colon into air-filled and fluid-filled segments. We used anatomical information in [5] to detect the AFB and its intensity was then set to that of air. Then all white structures (including fluid-filled colon parts) were set to the intensity of air as well. This resulted in setting the intensity of the bony areas (that have the same intensity of the fluid such as the spine, the ribs, and the pelvis) to the same intensity of the colon, to be all then discarded in the post processing stage. The main limitation of this cleansing procedure is that significant parts of the colon walls were missed, figure 1(b), along with the presence of gap-like artifacts, as they have the intensity between air and fluid which were neglected in this phase. This eventually led to missing polyps as will be shown later. The threshold applied above was that of the fluid (  ) which is high (255) and increases the chances of missing colon walls. If it is lowered to be between air and fluid in order to capture the walls, this will result in having non-colonic attachments that provide a lot of false positives and could not even be discarded in the post processing stage, figure 2 (b), (d). The proposed change to the pre-processing phase in [5] will thus be as follows: 1-Detect all connected components (clusters) in the volume, denoted as , along with the centroid  and the mean intensity   of each. 2014 Middle East Conference on Biomedical Engineering (MECBME) February 17-20, 2014, Hilton Hotel, Doha, Qatar U.S. Government work not protected by U.S. copyright 91