A novel computer aided detection (CADe) scheme for colonic polyps based on colon structure decomposition

To more accurately detect small polyps (ranged 5∼8mm, also called C2 stage) on the colon wall is of great significance for early diagnosis colorectal cancers. In order to achieve better detection results on small polyps in colon, both relatively global and local method were presented in the literature (or we can say single layer surface based or volume(image) based). However the colon usually consists of the Mucosa layers which result in Partial volume effect (PVE) on the colon wall, the task turns too complicated to be reached by simply following solo philosophy. According to the literature, the sensitivity of detection for small polyps only slightly above 0.83( polyp size range 5∼9mm). Though those results were based on different datasets due to share limitations, the average sensitivity is still not good enough for providing sufficient diagnosis support for further process. We believe that due to the influence of background undulating, small polyps and large polyps(bigger than 8mm) usually have not only some differences in geometric characteristics, but also Significant difference as far as the philosophy of measurement to be concerned. In order to adapt previous method to the mission of the small polyps' detection, a novelty global structure decomposition approach was suggested. By which the complex colon was separated into much uniform broken parts, where the Haustral folds and Haustral wall were broken into two independently kinds of morphological shape by means of analysis on second order derivatives of volume image. Polyp detection will look like finding certain abnormal protrusion on the relatively simple morphological surfaces. Hence the morphologically protruding characteristic of the small polyps will be enhanced in the volume image. Experimentally, we chose 60 patient cases from our database, and in which we focus on the polyps whose size range from 5∼8mm to validate the presented new approach. Compared with previously presented in the literature, the experimental are much more promising with average sensitivity of 0.99 per patient at mean time the false positive rate dramatically decreased to 1.2 per patient after false positive deduction based on texture.