An Efficient Colon Segmentation Technique Using a Two-stage Algorithm and Extracted Local Feature of Colonography Images

Colorectal cancer is one of the most widely occurring cancers among the aging population and the second leading cause of cancer related deaths. The use of 3D imaging to acquire virtual anatomic models of colon provides a minimally invasive diagnostic tool for the detection of colorectal polyps and cancer. This procedure termed virtual colonoscopy (V.C.) has in recent years been developed as an alternative method of massive population screening to examine the entire colon for early cancer detection. There are some problems which make colon segmentation a difficult task like the colon has complex shape and it is not the only gas-fill structure in abdomen, and a lot of areas with high CT number like bones, Contrast Enhanced Fluid (CEF) in colon and small bowel. A two-stage method for an automated segmentation of the colonic walls in volumetric CT data is proposed.

In this study, a two-stage algorithm was used. The aims of the first stage were to eliminate the extra objects and also to restrict the region of interest (ROI) in half-resolution as an input for fine process in second stage. In this phase, the multistage method performs segmentation using adaptive threshold, morphology operation and region growing using automated seed generation technique leading to a tissue characterization. This stage generates colon borders in half-resolution dataset. In the second phase, an edge sharpening, a modified region growing and an outer boundary tracking are applied to make fine borders in colon segmentation. The region growing in the second phase makes use of local adaptive thresholding instead of global view on database voxels.

To evaluate the results, certain tests for both stages of the algorithm were applied. In the total 82 datasets, the result showed 72 cases as excellent, 7 cases as good, 2 cases as fair and 1 case as poor in the segmentation process. Discussion and

An efficient segmentation method in half-sized volume data using an anatomical-oriented approach was developed to overcome the long processing time. The anatomical orientation guarantees the preservation of the segmentation accuracy while moving from half-resolution to full resolution. The output from the first stage can be used as a segmentation map for any complicated colon lumen extraction. The results demonstrate that the error of leakage to extra colonic can be minimized.