DETECTION AND SEGMENTATION OF COLONIC POLYPS ON HAUSTRAL FOLDS

Detections on haustral folds constitute a large portion of false positive findings in CT colonography, and polyps on the folds are more likely to be missed by a CAD system. This paper presents an approach to improve the segmentation of colonic polyps on haustral folds. The method is based on a combination of 3D knowledge-guided intensity adjustment, fuzzy clustering, and adaptive deformable model. We propose a dual-distance algorithm to detect the fold region. We then introduce a counter force in the model evolution to alleviate the over-segmentation problem that often occurs to polyps on haustral folds. The experiment was conducted on 395 patients with 83 polyps. The results were validated against manual measurement. The volumetric measurements were strongly correlated and there was no significant difference (P = 0.37 in paired t-test). The median Dice coefficient for volume overlap between automatic and manual segmentation was 0.75 (standard deviation 0.15). The counter force improves the segmentation accuracy of polyps on-fold by 21%.