| According to up-to-date statistics from the American Cancer Society, colorectalcarcinoma is the third most common cancer in both men and women worldwide. Theincidence of colon and rectal cancer was decreasing in the past twenty in America. Thisshould be highly attributed to the regular colonoscopy screening examinations forpopulation over50that were recommended by health institutions. Nevertheless, Incontrast to the overall decline, among younger adults less than50years old, who are not atthe average risk and the screening is not recommended for, colorectal cancer incidencerate has been increasing by1.6%per year since1998. American Cancer Society hasproved that the regular colonoscopy screening examinations by a healthcare professionalcan result in the detection and removal of precancerous polyps at an early stage when thegrowths are most treatable. One of the critical reasons for the increasing of the incidencefor people younger than50is that most of the people in this population don’t have thehabit of regular colonoscopy screening examination, for the rather complicatedpreparation procedure and intolerable examine process of optical colonoscopy. Many of them missed the opportunity of finding the lesions at early time, which is a golden stagefor treating before they developed into cancer.Data from the Ministry of Health (National health and family planning commission)shows colon and rectal cancer is the5thmost mortality rate among all the cancers in theyears from2004~2005. The mortality rate increased by36.8%and57.6%when comparedto1990~1992and1973~1975respectively. The colorectal cancer has been posed a hugethreat to the health of people in our country. The dramatically increased mortality rate isworth of close attention. The increased mortality rate may be caused by two reasons:1)the increasing incidence of colorectal disease that accompanied with the changing ofdietary habit and2) the relative lagged improvement of medical environment whencomparing to the rapidly changing economy and health consciousness.According to many researchers’ reports, it takes about5~15years for a colon lesion toevolve into malignant. One of the most characters of colon lesion is that it has no obvioussymptoms in its early age. Clinical practices have proved that a regular colon screeningexamination in every5years can significantly reduce the incidence of colorectal cancer.Nevertheless, the optical colonoscopy, currently the gold standard of colon lesionexamination and invasive, is hard to be applied in the situation of regular screening for itrequires a rather complicated pre-process preparation and high in-process patientendurance. This resulted in the current situation that many patients missed the gold time todetection the lesion and treating it in its early age. Clinical data shows that the5-yearsurvive rate can reach90%if colon lesions are detected and treated in their early age. Sothe early detection and diagnosis of colon lesions play a very important role in the treatingprocess.Virtual colonoscopy (VC) provides an alternative for colon screening. It has thecharacteristics of minimal invasive, reduced examine time and improved patient testingexperience. VC is expected to be a novel colorectal screening method for mass people forits many tangible advantages. Nevertheless, there are still some challenges in the domainof VC research regardless some critical progress in the past years. The modern CT canprovide a lot of data in the thin slice scan mode. Traditional VC needs a doctor to browse the data in3D and2D to locate lesions and the long time browsing can cause fatigue,inefficiency and even missing some lesion. As the developing of CADe technique, recentyears some researchers began to use this technique to help doctor to locate lesions.Usually CADe system relies on morphological changes in exception to identify lesions forthere is no color information in CT scans, so there have many false positives for left feces,folds of colon etc. always mimic the shape of polyps. This false positive still make theexamine procedure less effective. Another problem is even a lesion candidate is presentedto the doctor, it’s still difficult to differentiate for CT data don’t provide information suchas color and texture that are always provided by optical colonoscopy. All these problemsimpeded the applying of CADe and the developing of CADx.To resolve the problems mentioned above, we researched some key techniques ofcolon CAD in this study. Besides the non-invasive and easy to implement traits, theobtained CT slices can be deemed as a volumetric data. We proposed a3D textural modelbased on the suspicious lesion patches and evaluated the possibility of applying them inCADe of colon to reduce false positive rate. Considering the essence of the proposedmodels are the internal patterns of CT value of a lesion, something like clinicalpathological testing (magnifying the lesion, observing the texture and judging if a lesion isbenign or malignant), we also explored the applying of the proposed3D textural model inthe domain of CADx.The evaluating data are from Prof. Perry Pickhardt’s group (Wisconsin University),which includes67patients and each of them has undergone both supine and pronepositions scan which yield134scan sets in total. There are95polyps, sized from4mm to30mm, were found in both optical colonoscopy and CT colonoscopy. In order to get moresamples, in our study we treat two different scans of one patient as two cases, and thus wehave190scan cases. To improve the test experience, we also implemented the electroniccleansing technique in the study, all the patients only required to eat fluid food with somecontrast but the laxative is no longer needed. Based on these data, what we have done areas follows. 1. Create a3D model for texture analysisThe purpose of the textural models is to extract the internal CT value pattern to formsome new features and use them to differentiate the VOIs. Unlike the traditional features,the proposed3D textural models can take full use of the voxel information in the volumeand doesn’t strongly depend on the segmentation quality of lesion, thus can provide atotally new viewpoint for the research of colon CADe and CADx.The proposed textural models can be established in two steps:1) compute gray levelco-occurrence matrices (GLCM) and gray level and gradient co-occurrence matrix(GLGCM) and2) extract some statistical information from these matrices. The GLCMreflects the co-occurrence relationships of the voxels within the volume of interest. TheGLGCM extracts the relationships between the VOI and its gradient volume by means ofcounting the value co-occurrences of the two volumes. Finally some statistical features areextracted from these matrices and used to differentiate VOIs. These co-occurrencematrices reflect the internal CT pattern and can render some details in contrast to meanand variance which only presents the global features of VOIs.In specificity, the proposed GLCM model extracted26GLCMs along26directionswhich uniformly distributed on a spherical surface and the GLGCM model computed1GLGCM from the image and its corresponding gradient image. From each GLCMs13Haralick were computed. To obtain the isotropic trait, for each feature the statistical valuesof mean and range (maximum minus minimum) were computed along26directions. The13means and13ranges of one VOI form a26-feature GLCM feature vector. As we doneon a single GLCM,13Haralick features were computed from GLGCM, plus the26GLCM features finally we have39volumetric textural features in total.2. Evaluate3D textural models CADe of CT colonoscopyTo evaluate whether the proposed models can be used in CADe of CT colonoscopy,we used the models to analyze the manually outlined polyp volumes and normal tissuevolumes that were mixed together and computed the specificity and sensitivity. Also weapplied the models in a previously developed CADe pipeline to test the possibility of practical use and their performance.In the manual scheme we called the volume of lesions VOI and volume of normaltissue VON respectively. From the190polyp cases we outlined190VOIs. For every VOIa VON was outlined nearby under the supervising of an experienced radiologist. Based onthe3D textural models,26GLCMs (abbreviated as C26) were extracted from these VOIsand VONs.13textural features were computed from GLGCM (abbr. G13). Besides, wealso computed6traditional features (abbr. T6), i.e. mean and variance of volume CTvalues, mean and variance of shape index (SI), mean and variance of curvedness (CV).In the automatic scheme,3211initial polyp candidates were extracted by the CADepipeline, which include94true polyps and1missed4mm polyp. There19polyps weredetected in either of the two scan positions for the interference of partial volume effect ofcontrast. The number of detected polyp cases is169. So the by polyp sensitivity is98.94%and by case sensitivity is88.95%.The well-known support vector machine (SVM) with radial basis function kernelwas applied to evaluate the performance of the proposed texture features using manuallyoutlined or automatically clustered IPCs. Six different combinations of these features, i.e.feature vectors, were tested during the evaluation process. They are1)6traditionalfeatures (T6),2)26GLCM-derived features (C26),3)13GLGCM-derived features (G13),4)6traditional features plus26GLCM-derived features (T6+C26),5)6traditionalfeatures plus13GLGCM-derived features (T6+G13), and6)6traditional features plusboth GLCM and GLGCM-derived features (T6+C26+G13). For feature vectors includingC26, there were3versions corresponding to different distance d of1,2and3respectively.According to SVM rule, we need to split the dataset into training group and testinggroup. Our grouping scheme is: half of feature vectors from TPs and FPs were randomlychosen, respectively, and mixed to form the training set (In the manual scheme, VOIs weretreated as TPs and VONs as FPs). The left feature vectors were grouped as the test set. Thetwo-fold cross-validation strategy was used in the classification procedure. That is, foreach training and test sets randomly grouped, in the1st fold the SVM was optimized andtrained by the training set, followed by the testing on the test set. Then the two data sets were exchanged in the2nd fold and all the parameter searching, training and testingprocess was performed again. The performance was measured by the averaged area underthe receiver operating characteristic curve (AUC) based on the testing results. In this study,the random grouping and the two-fold cross-validation were repeated100times for eachfeature combination.In the manual scheme, the AUCs were (98.17±0.97)%,(95.53±1.12)%and(97.63±0.79)%if T6, G13and C26(d=1) were used respectively. The AUCs were(98.49±0.78)%and (98.89±0.59)%when feature vectors were T6+G13and T6+C26, i.e.G13and C26were used as supplementary features. We can see both the traditionalfeatures and proposed volumetric textural have good differentiating ability in the manualscheme and the performance was slightly improved when the volumetric textural featureswere supplemented to T6. In the automatic scheme, AUCs were (83.69±5.93)%、(70.22±3.58)%、(73.84±3.18)%、(89.46±2.45)%and (90.34±2.11)%for feature vectors ofT6, G13, CT6(d=1), T6+G13and T6+C26(d=1) respectively. When G13and C26wereused as supplementaries, AUCs were improved by5.77%and6.18%respectively. Butwhen the vector T6+G13+C26was used, no significant improvement was seen with theAUC of (98.74±0.66)%and (90.06±2.36)%in manual scheme and automatic pipelinerespectively.3. Explore the possibility of CADx in CT colonoscopyBased on the manually outlined VOI, all the VOIs were classified into5pathologicalcategories according the pathological test. They are1)190normal cases,2)56hyperplastic cases,3)94tubular adenoma cases,4)34tubuvillous adenoma cases and5)6carcinoadenoma cases. Six principal features were chose according to a componentanalysis (PCA) after the volumetric textural features were computed. The HotellingT-square test showed that when the significant probability was set to P <0.05, significantdifferences were found between any arbitrary paired groups of the5pathological typesexcept the pair of hyperplastic group and tubular adenoma group.To further explore the effect of applying the textural features in CADx in CT colonoscopy, we tried to further differentiate the true positives, which were detected by thepreviously developed CAD pipeline. We tried to classify the IPCs into benign andmalignant after we made the standard of benign and malignant for the4pathologicallesions in3ways. The classifiers used in CADx were SVM and Random Forest (RF). Likewhat we done in CADe part, ROC curves were plotted to evaluate the performance.The output of CADe presented169positive lesions, sized over5mm. According topathological report, there are50hyperplastics,82tubular adenomas,31tubuvillousadenomas and6carcinoadenomas. To include more features in our study, we extracted10geometrical features and64textural features, many of which were proposed by otherresearchers or our previous study. AUCs can reach89.78%and85.2%for SVM and RFrespectively when hyperplastics and tubular adenomas were grouped as benigns,tubuvillous adenomas and carcinoadenomas were grouped as malignants. AUCs reached86.16%(SVM) and3.62%(RF) when carcinoadenomas were grouped as malignants andothers grouped as benigns.Summary:In this study we extended the traditional2D Haralick textural model to3D andfurther optimized it to a non-parameter model. The proposed3D textural models weredesigned to reflect the internal patterns of the analyzed volume. They have the traits ofisotropic, taking full use of the voxels of the volume and less depending on segmentation.The evaluation and testing in the CADe of CT colonoscopy showed that the improvementswere not obvious when the proposed textural features were used alone. But if the proposedtextural features were used as supplementaries to traditional features, the differentialability of the CADe system was significantly improved. We also explored the possibilityof applying3D textural in CADx of CT colonoscopy. Statistical test showed significantdifference between the proposed textural features of different lesion types. The proposed3D textural features can yield AUC over85%when be used to classify benigns andmalignants on clinical data. In the grouping scheme that hyperplastics and tubularadenomas were labelled as benigns and tubuvillous adenomas and carcinoadenomas labeled as malignants, the AUC from SVM can reach89.78%. These preliminary resultsshowed the potential and value of applying the proposed3D textural features in CADx ofCT colonoscopy.The innovation points of the study includes:1) optimization of the3D texturalmodels,2) evaluating the application of the proposed3D textural models in CADe of CTcolonoscopy and3) for the first time explored the possibility of applying these models inCADx of CT colonoscopy and providing critical technique for electronic biopsy. |
PDF Full Text Request