Texture Analysis And Detection Of Polyps From Colon Wall For Virtual Colonoscopy

Colorectal cancer is one of the most common malignant tumors.According to statistics of the United States authorities in 2018,the incidence and mortality of colorectal cancer in the United States are in top three among all cancers.Most colorectal cancers begin as a polyp,early detection of polyps by screening and timely treatments are effective ways to reduce the incidence and mortality of colorectal cancer.Compared with colonoscopy and other traditional tools,virtual colonoscopy(VC)is minimally invasive,less time-consuming,easily tolerated and especially suitable for high-risk large population based screening.The computer-aided detection of polyps based on virtual colonoscopy assists the doctors to detect the lesions,which can reduce the workload and increase the detection accuracy effectively.But most traditional computer-aided detection methods are based on the accurate segmentation of the inner colon wall,detecting polyps by shape variation parameters such as the shape index,curvature and so on,their common defect is that only focus on the shape difference of the inner surface and neglected the nature of the tissue between inner and outer colon wall.As a result,non-polypoid lesions which may have a higher risk,such as flat polyps and small sessile polyps are difficult to be detected.In fact,the atypia of tumor tissue causes tissue density changes in the polyp and its attenuation coefficient of X-ray will change accordingly,which can be reflected in the CT images,but these differences are subtle,and present inside the polyp which are hard to notice by the naked eye.However,three-dimensional texture features can be quantified by the gray-level distribution and permutation characteristics to reflect these differences.At the same time,the intestinal tagging agent is widely used in the virtual colonoscopy examination,which is often attached to the lesion and cause specific texture changes.To evaluate the performance of three-dimensional texture features in differentiation between polyps and normal colon wall tissue and apply them to the automatic detection of colonic polyps,texture analysis and detection of polyps from colon wall for virtual colonoscopy were carried out,mainly including the following two aspects of the work.(1)Texture analysis on colorectal polyps based on virtual colonoscopy.Volume of interest(VOI)was delineated from colorectal polyps and normal wall tissue,and gradient and curvature maps were calculated on these VOI,totally 198 features were derived from each VOI.The optimal feature subset of 48 features were selected as by feature selection method,and they were input into four different classifiers to differentiate polyps from wall tissues,the average sensitivity were all over 0.99,the average specificity were all over 0.98,the average accuracy is 0.97 and the average AUC was all up to 0.99,indicating that the optimal feature subset can effectively reflect the differences between polyps and normal wall tissues in the colon,and can be used as a good marker to characterize the atypia of polyps in the colon,and provide the basis for the detection of colon polyps based on texture analysis in the follow-up study.At the same time,research on differentiation of polyps of different shapes from normal wall tissues,and differentiation of benign and malignant polyps were also carried out.In differentiation of benign and malignant polyps,the average sensitivity were all over 0.99,the average specificity were all over 0.98,the average accuracy is 0.98 and the average AUC was all up to 0.99;In differentiation of polyps of different shapes from normal wall tissues,the average sensitivity were all over 0.76,the average specificity were all over 0.81,the average accuracy is 0.79 and the average AUC was all up to 0.89.Through experiments,we found that the extracted texture features not only have very high classification performance for the differentiation of normal wall tissues and polyps,but also can effectively distinguish between benign and malignant polyps,which further corroborated our above-mentioned theory.(2)Colorectal polyp detection based on texture features.Full colon model containing outer and inner colonic wall were obtained by segmentation,and the skeleton that could represent the topological structure of the whole colon was calculated according to the segmentation results,fixed size neighborhood were taken for each voxel of the skeleton,taken the spatially intersection regions of the full colon model and neighborhood as VOI,texture features were extracted on each VOI,and then initial polyp regions were obtained by anomaly detection method based on optimal feature subset.Although the above process has been screened for initial polyp regions,the false positives were very high,actions were taken to further reduce false positives such as cutting the colon with offsets and connected region merging,finally,judgment on polyps is made through random forests.Our detection method were validate based on 20 sets of virtual colonoscopy image data with 36 polyp,including 16 flat polyps,18 sessile polyps,and 2 pedunculated polyps,the experiment results showed that the proposed detection method can achieve 80% detection sensitivity,and the number of false positives per dataset was 9.98.Proposed method in this paper broke through the traditional computer-aided detection method of colonic polyps which relying only on the shape of polyps so that missed most of flat polyps and small sessile polyps,but these non-polypoid lesions can be effectively detected by our new workflow.Complementing with traditional methods and further reducing false positives are expected,this will effectively improve the overall performance of computer-aided detection on colonic polyps,promote the application of virtual colonoscopy in clinical colorectal cancer screening.