Detection Of Small/Flat Polyps Based On Intact Colon Wall And Texture Features

According to the data released by the American Cancer Society(ACS)in 2020,the incidence rate and mortality rate of colorectal cancer(CRC)ranked the third among all cancers,and it ranked fifth in China.Colorectal cancer is the malignant transformation of colorectal polyps in 5-15 years.Timely finding and removing polyps through screening is the most effective way to reduce the incidence rate and mortality rate.Compared with optical colonoscopy(OC),virtual colonoscopy(VC)has the advantages of painless,noninvasive,fast and effective,safe use,etc.,which is more willing to be accepted by patients.The technology of computer-aided detection(CADe)based on VC can help doctors to detect the position of polyp automatically,improve the detection efficiency and avoid the mistakes brought by subjective experience.However,the traditional polyp CADe technology mainly relies on the morphological changes of the inner wall of the colon to detect,the small polyp and flat polyp with insignificant shape change are easy to be overlooked.The average detection sensitivity of some studies for 6-9mm small polyps was 72%,and the corresponding false positive number was 8.12/dataset,but the detection sensitivity of flat polyps was only 46%.Considering that flat polyps have a higher risk of canceration,and the detection of 6-9mm small polyps is helpful to the early detection of rectal cancer,more effective auxiliary detection methods are urgently needed to improve the detection ability of small and flat polyps.Studies have shown that there are different degrees of differences in cell morphology and tissue structure between tumor and its original normal tissue,which can be shown by CT value distribution and texture changes.Most of the researches on colon polyps detection use the texture features to reduce false positives.Considering that all colonic polyps originate from the colon wall,if we can effectively extract the complete colon wall and combine with the identification ability of texture features,can we improve the detection performance of CADe for small / flat polyps? The CADe technology of colorectal polyp based on complete colon wall and texture features is carried out to verify this idea in this paper.The research contents mainly include the following three aspects:(1)Detection of suspected colorectal polyps based on complete colon wall and texture features.The framework mainly includes four key parts: 1)The acquisition of the whole colon wall: in order to obtain the area for texture feature extraction,the expectation maximization algorithm of maximum posterior estimation is used to segment the inner wall of the colon and obtain the single voxel contour representing the topology of the colon,and then the outer wall of the colon is segmented combined with the level set algorithm.The inner and outer walls and inter wall tissues of the whole colon are obtained.2)The acquisition of volume of interest(VOI): a spherical sliding window is used based on the acquired colon wall to slide on the colon contour point by point,and the area overlapped with the inner and outer walls of the colon is taken as the VOI for texture feature calculation.3)Texture feature extraction: totally 267 texture features are extracted from each VOI,including 18 histogram features extracted from CT image,gradient image and curvature image and 180 texture features extracted from their gray level co-occurrence matrix and 69 features extracted based on gray level run length matrix(GLRLM),gray level size zone matrix(GLSZM),neighborhood gray tone difference matrix(NGTDM),grey level distance zone matrix(GLDZM),neighbouring grey level dependence matrix(NGLDM).4)Suspected polyp detection based on classifier: Ada Boost classifier is used for classification and prediction,and the initial suspected polyps are obtained by regional fusion of prediction results.50 sets of training sets(including 89 polyps)and 20 sets of test sets(including 58 polyps,including 14 flat and 52 6-9mm)are used to test the proposed framework.The results showed that the detection sensitivity of the method was100%,but the corresponding false positive number was as high as 373.75/dataset.Although the false positives are high,all polyps can be detected completely,which indicates that it is feasible to detect small / flat polyps based on complete colon wall and texture features.(2)Optimization of algorithm for detecting suspected colorectal polyps based on texture featuresIn order to improve the performance of the algorithm,the above detection algorithm is optimized,mainly including three aspects: 1)Optimization of sliding window selection:the size of sliding window has a great impact on the detection performance of the system.In this study,the VOI is selected by spherical sliding windows with diameters of5,7,9,11,13,15.Then the results are evaluated according to the specificity and false positives.The results show that the best performance is obtained when the diameter of sliding window is 13;2)Optimal selection of classifiers: the performance of Ada Boost classifier needs to be further verified.Firstly,based on the mean decrease integrity(MDI)algorithm of random forest,the optimal feature subset containing 102 features is selected.Then,,the random forest classifier is used for prediction based on the optimal feature subset,and the prediction results are compared with Ada Boost classifier.The results show that Ada Boost classifier has better performance.3)Optimization of training set selection:because the samples are extremely unbalanced,it is difficult to directly use for learning.Therefore,in this study,the false positive samples are added to the training set by iteration,so as to learn the texture features of normal colon tissue more comprehensively.Through the above optimization,when the detection sensitivity is 100%,the corresponding false positive number is 40.25/dataset.The experimental results show that the detection performance of the system can be effectively improved by optimizing the algorithm parameters and training set.(3)Reducing false positives based on sparse preserving projection model and multi feature combinationIn order to further reduce the false positives of detect results and improve the detection performance of the system,three studies are carries out in this part,mainly including: 1)Sparsity preserving projections(SPP)classification model: in order to reduce false positives by utilizing the gray difference between polyps and colon tissues,we use the sparse reconstruction relationship between SPP preserved data and the sparse representation based classification(SRC)is used for classification.When the detection sensitivity is 100%,the number of false positive is further reduced to 29.3/dataset..2)Morphological feature threshold model: firstly,five morphological features(including wall thickness,shape index,curvature,curvature,curvature,and curvature difference between inner and outer walls)are calculated respectively.Then,the maximum and minimum characteristic values of all voxels in each polyp are calculated to obtain the range of all voxel characteristic values of all real polyps as the threshold.If the characteristic values of all voxels of suspected polyps are beyond the threshold,they will be excluded as false positive.When the detection sensitivity is 100%,the number of false positive is further reduced to 15.35/dataset..3)The highest prediction score model: firstly,the domain tissue of each voxel of suspected polyps based on the variable ellipsoid,and then the highest predictive score of all voxels of suspected polyp is taken as its prediction score,and the false positive is reduced by increasing the screening threshold.When the detection sensitivity is 100%,the number of false positive is further reduced to29.3/dataset.The experimental results show that the three methods can reduce false positives while maintaining sensitivity.The method proposed in this study overcomes the problem of poor detection ability of traditional CADe for small / flat polyps by using morphological changes of colon inner wall,and the automatic detection of small and flat polyps is effectively realized.Furthermore,the number of false positives is greatly reduced and the performance of CADe based on VC is effectively improved through the optimization of algorithm parameters and training set,as well as sparse preserving projection model and the fusion of morphological features.