| BackgroundIn China, colorectal cancer （CRC） is a common malignant tumor in the digestive system. The morbidity of CRC is located in the fifth place of malignant tumor, and the mortality rate ranks fifth in the cancer related death. Colorectal liver metastasis is the most common distant metastases from CRC, which is also the main cause of death for CRC patients. Up to 15%-25% of patients present with synchronous colorectal liver metastases （sCRLMs） at the time of diagnosis of their primary CRC. There was a significant difference in the treatment modality between the colorectal cancer patients with synchronous colorectal liver metastasis and patients without liver metastasis. Given surgical resection of all tumor sites and postoperative comprehensive treatment, sCRLMs patients’5-year survival rate was 62%. Therefore, the presence of synchronous liver metastases is the key factor affecting the treatment decision and the detection of synchronous liver metastases provides an essential basis for individualized treatment plan. At present, the widely used method for the diagnosis of synchronous colorectal liver metastases is through the detection of carcinoembryonic antigen （CEA）. However, the CEA is not a specific antigen for CRC, clinical practice has demonstrated the lack of it in specificity. Indicator of colorectal liver metastasis currently reported in the literature including the primary tumor pathological grade, histological type, lymph node metastasis, invasion of the intestinal wall, and tumor marker. However, these indicators are invasive and only postoperatively available. Therefore, there is still lack of a convenient biomarker that can be obtained preoperatively for the detection of synchronous colorectal liver metastases.CT （Computed Tomography）, as a non-invasive imaging tool, is the first choice for the preoperative evaluation of colorectal cancer patients. CT has been widely used in the diagnosis of colorectal cancer liver metastasis to provide an objective and comprehensive description of liver metastasis. Unenhanced CT examination can be used to describe the overall property of the liver lesions （solid or cystic; with or withour calcification）; while contrast-enhanced CT examination can further reflect the blood supply of the liver lesion. However, simply based on the primary foci to predict distant metastasis （M stage） is beyond the capacity of traditional CT imaging evaluation （either unenhanced CT or contrast-enhanced CT）. Meanwhile, lesions smaller than 10mm in diameter are difficult to be detected and tend to show similar charateristics to small hepatoangioma and liver cysts. In addition, routine analysis of the CT image can be used to assess the structural characteristics of the tumor, but fails to describe the details of the function or the molecular property of the tumor.In recent years, the rapid development of computer hardware and software technology has greatly promoted the application of texture analysis in medical image. Texture analysis is a technique to evaluate the focal feature of pixel gray value, along with its variation and distribution pattern. Different texture features extracted using different methods can reflect physiological or pathological information from different aspect. Compared with the traditional CT imaging evaluation, which simply describe qualitative characteristics of the region of interest （ROI） （such as lobulated tumor shape, boundary）, texture analysis can provide a more comprehensive detection of the physiological heterogeneity within the region of interest, which can be applied to a variety of medical images to assist in the diagnosis and treatment of diseases. Extracted texture features are "measurement and quantifiable, can be as healthy or physiologically relevant conditions (such as disease risk assessment, disease diagnosis, metabolic process of evaluation indexes of biological parameters", which thus have the potential to serve as imaging biomarkers. Gray histogram and gray level co-occurrence matrix （Grey-level Co-occurrence matrix, GLCM） texture features extracted from CT images have been reported to have association with tumor staging, metabolism, hypoxia and angiogenesis, which have also shown potential in detection and characterization of tumor, and the assessment of curative effect. There have been several studies conducted with the texture analysis of CT images in colorectal cancer. It was found that the texture analysis of liver parenchyma extracted from portal vein phase CT image showed better performance in predicting survival of colorectal cancer patients compared with that of the CT perfusion image, suggesting that the different prognosis CRC patients can be reflected in the texture features of liver CT image. Conversely, whether can the the imaging biomarkers established with texture features extracted from the primary colorectal cancer lesion reflect change in liver and detect synchronous colorectal liver metastases, is worthy of exploring.Purposes1. To establish an imaging biomarker to detect synchronous colorectal liver metastases based on the CT texture features of colorectal cancer.2. To establish a predictive model integrating the imaging biomarkers for the detection of synchronous colorectal liver metastases.Materials and MethodsStudy populationEthical approval was obtained for this retrospective analysis by our institutional review board, with informed consent waved. Consecutive patients confirmed with colorectal adenocarcinomas from January 2007 and April 2010 were reviewed to form the cohort of this study.The inclusion criteria and exclusion criteria of the LM-positive cohortThe inclusion criteria for this study were as follows:（a） patients who were confirmed to have synchronous colorectal liver metastases when diagnosed; （b） standard unenhanced and contrast-enhanced CT performed within two weeks before or after the diagnosis of CRC.The exclusion criteria for this study were as follows:（a） patients were diagnosed to have liver metastasis from other primary cancer, （b） patients underwent therapy （radiotherapy, chemotherapy or chemoradiotherapy） before the base line CT examination; （b） inflammatory diseases, including infections, ischemic heart disease, collagen diseases, bowel perforation or obstruction; （c） familial adenomatous polyposis or hereditary non-polyposis colon cancer; （e） poor imaging quality not quantified for image analysis.The inclusion criteria and exclusion criteria of the LM-negative cohortThe inclusion criteria for this cohort were as follows:（a） patients who were confirmed to have CRC and stayed free of colorectal liver metastases when diagnosed; （b） standard unenhanced and contrast-enhanced CT performed within two weeks before or after the diagnosis of CRC.The exclusion criteria for this study were as follows:（a） patients underwent therapy （radiotherapy, chemotherapy or chemoradiotherapy） before the base line CT examination; （b） inflammatory diseases, including infections, ischemic heart disease, collagen diseases, bowel perforation or obstruction; （c） familial adenomatous polyposis or hereditary non-polyposis colon cancer; （e） poor imaging quality not quantified for image analysis.Baseline data collectionBaseline data pertaining to demographics, primary tumor site, and carcinoembryonic antigen （CEA） level were reviewed and recorded. Note that the threshold value for the level of CEA was≤ 5ng/ml and> 5ng/ml according to the normal range used in our institution.CT Imaging ProtocolAll patients underwent contrast-enhanced abdominal CT using one of the two multi-detector row CT （MDCT） systems （GE Lightspeed Ultra 8, GE Healthcare, Hino, Japan or 64-slice LightSpeed VCT, GE Medical systems, Milwaukee, Wis）. The acquisition parameters are as follows:120 kV; 160 mAs; 0.5-or 0.4-second rotation time; detector collimation:8×2.5 mm or 64x0.625mm; field of view,350×350 mm; matrix,512x512. After routine non-enhanced CT, arterial and portal venous-phase contrast-enhanced CT were performed after 22s and 60 s delay following intravenous administration of 90-100 ml of iodinated contrast material （Ultravist 370, Bayer Schering Pharma, Berlin, Germany） at a rate of 3.0 or 3.5ml/s with a pump injector （Ulrich CT Plus 150, Ulrich Medical, Ulm, Germany）. Contrast-enhanced CT was reconstructed with reconstruction thickness of 2.5 mm. Image Post-processing and Texture AnalysisPortal venous-phase CT images （thickness:2.5 mm） were CT images were retrieved from the picture archiving and communication system （PACS） （Carestream, Canada） and transferred to a Workstation （Viewforum; iMAC; Cupertino, CA, USA） for post-processing. Texture analysis was applied to the CT images using in-house texture analysis software with algorithms implemented in Matlab 2010a （Mathworks, Natick, USA） by two radiologists independently （Reader 1 and Reader 2 with 5 and 3 years of clinical experience in abdominal CT, respectively）. Reader 1 extracted the texture feature from CT images twice in a 2-week interval following the same procedure to assess the intra-observer reproducibility; whose first measurement was used to compare with Reader 2’s extraction to assess the inter-observer agreement.ROI （region of interest） settingAn ROI was delineated initially around the tumor outline for the largest cross-sectional area on portal venous phase CT image. ROIs on the unenhanced CT and arterial phase CT image were delineated with reference to that on the portal venous phase CT image. The ROI was further refined by excluding air area with a thresholding procedure that removed from analysis any pixels with attenuation values below-50 HU and beyond 300 HU.Extraction of texture featuresA Laplacian of Gaussian spatial band-pass filter （V2G） was used to derive image features at different spatial scales by turning the filter parameter between 1.0 and 2.5 （1.0,1.5,2.0,2.5）.150 imaging texture features from the category of histogram and gray-level co-occurrence matrix （GLCM） were extracted from one single CT image. In total,450 texture features were ultimately extracted from the unenhanced CT image, arterial phase image and portal venous phase image.Statistical AnalysisDemographics characteristicsThe age of LM-positive and LM-negative patients was firstly tested with the Kolmogorov-Smirnov test for normality and Levene’s test for variance homogeneity, the difference between which was then accessed using independent t-test or Manney-U test. To assess the difference in categorical variables （gender, primary tumor site, and CEA level） between LM-positive and LM-negative patients, Chi-Squared tests were used.The intra-observer and inter-observer agreementThe intra-observer and inter-observer agreement of feature extraction were evaluated by using intra-class and inter-class correlation coefficient （ICC）. Intra-observer ICC was computed from Reader 1’s two extraction. Inter-observer ICC was computed from Reader l’s first extraction and Reader 2’s extraction. An ICC greater than 0.75 was considered good agreement. The intra-observer and inter-observer agreement were evaluated for the extraction on unenhanced CT image, arterial phase CT image and portal venous phase CT image, respectively.Feature selection and imaging biomarker buildingThe least absolute shrinkage and selection operator method （Lasso）, which is suitable for the regression of high dimensional data, was used to select the most useful prognostic features. Three imaging biomarkers were built through a linear combination of the selected features weighted by their respective coefficients, for the unenhanced CT image, arterial phase CT image and portal venous phase CT image. Three hazard scores were calculated for each patient.Evaluation of the performance of imaging biomarker for the dection of sCRLMsThe difference in the biomarker-based hazard scores between the LM-positive and LM-negative patients was assessed using independent t-test or Mann-Whitney U test.Receiver operator characteristic curves （ROCs） were performed to assess the performance of imaging biomarkers for the detection of sCRLMs. Area under the curve, sensitivity, specificity, positive prediction value and negative prediction value were recorded, each including the 95% confidence interval. The cut-off points of the ROCs were determinded by the maximum Youden index （sensitivity+specificity-1）.Validation of the imaging biomarker for the dection of sCRLMsStratified analyses were conducted using Mann-Whitney U test or independent t-test to explore the potential association of each of the imaging biomarker with the LM status within subgroups of clinical risk factors （age; gender; primary tumor site; CEA level）.Multivariable logistic regression analysis starting with the following clinical risk factors （age, gender, primary tumor site, and CEA level） and the imaging biomarkers was applied to develop a diagnostic model for the sCRLMs. Backward stepwise selection was applied, using the likelihood ratio test with the stopping rule being Akaike’s information criterion （AIC）.Statistical softwareThe statistical analysis was conducted with R software （version3.0.1; http://www.Rproject.org）. The reported statistical significance levels were all two-sided, with statistical significance set at 0.05. Lasso binary logistic regression was done using the "glmnet" package. Multivariate binary logistic regression was done with the "rms" package. ROC analysis was performed the "Hmisc" package.ResultPopulation demographicsLM-positive cohort:eighty LM-positive patients were enrolled in this study.62 men and 18 women; mean age,63.24 years ±12.57 （standard deviation）, age range, 24-85 years; mean age of men,62.13 years ± 12.01, age range 24-80 years; mean age of women 65.13 years ± 12.66, age range 26-85 years.LM-negative cohort:eighty LM-negative patients were enrolled in this study. (62 men and 18 women; mean age,63.24 years ±12.57 （standard deviation）, age range, 24-85 years; mean age of men,62.13 years ± 12.01, age range 24-80 years; mean age of women 65.13 years ± 12.66, age range 26-85 years.There was no significant difference in the age between the LM-positive and LM-negative cohorts.Inter-observer and intra-observer reproducibility of radiomics feature extractionThe intra-observer ICC calculated based on Reader l’s two measurements ranged from 0.821 to 0.977. The lowest ICC 0.821 was from unenhanced CT image, and the highest ICC 0.977 was from portal venous phase CT image. Inter-observer agreement between Reader l’s first measurements and Reader 2’s measurements was good for either the unenhanced CT image, arterial phase CT image or the portal venous phase CT image, with ICCs ranging from 0.795 to 0.933. For the inter-observer agreement, the lowest ICC 0.795 was from unenhanced CT image, and the highest ICC 0.933 was from portal venous phase CT image.Constructed imaging biomarkerA hazard score was calculated for each patient through a linear combination of selected features weighted by their respective coefficients, with respect to each of the imaging biomarkers.Constructed imaging biomarker based on unenhanced CT image:Hazard Score=-3.635+0.007 × his50mean0+0.189 × contrast00-1.7450 × skewness1.0+0.03550 × his50mean1.0 +2.555 × homogeneiy901.0+0.743×correlation1351.5 +1.215 × correlation1352.5Constructed imaging biomarker based on arterial phase CT image:Hazard Score=-22.938+6.053 × entropy00+0.023×his50mean0 +0.1400 × his10mean0-1.1560 × skewness1.0-0.211×homogeneiy451.0+4.902×homogeneiy1351.0-0.003×his 10 SD 1.0-1.635×correlation 45 2.0 +3.711×homogeneiy1352.0+2.682×correlation902.5 +0.888×correlation1352.5+0.0140 × his50mean2.5 +0.001×his25mean2.5Constructed imaging biomarker based on portal venous phase CT image:Hazard Score= 3.351-0.0531 × kurtosis0+0.0360 × contrast00 +0.001 × contrast450+116.885×energy00 +0.0210 × hismean0+2.35400×skewness1.0-0.971 ×kurtosis1.0+0.115×contrast451.0 +7.303 × homogeneiy901.0+0.001×hismean2.5 +0.0141 ×hisean1.0+0.6 x hOmogeneiy1351.0-11.257 × homogeneiy01.5+0.002×hismean1.5 +0.0565×his10mean1.5+0.0134×hiS10mean2.0 +0.001×hiSmea2.0-4.530×homogeneiy452.0-0.252×contrast1352.5-8.929×homogeneiy02.5 +0.920 ×homogeneiy902.5-0.0700×kurtosis1.5 +0.009×hiS25mean2.54.Constructed combined imaging biomarker based on CT image:Hazard Score=-63.125+0.072×ucontrast00+0.821×uentropy450 0.041×uhis10SD00.3870×u-homogeneiy-450-4.0440×uskewness1.0+109.0760×uenerg901.0 +0.088×uhomogeneiy901.0+0.875×uentropy451.0 +0.036×uhis50mean1.0+0.527×ucorrelation1352.5-0.011×uhis.25mean2.54.5610×ahomogeneiy450 +8.2430×aentropy00+1.5940×acorrelation1352.5 +0.26600×ahis10..mean0-3.20100×askewness1.0 +30.9200×aenergy90.1.0+0.0080×ahis50mean-1.0 +0.0340×ahis25mean1.0-0.0660×ahis10SD1.0-1.036×acorrelatioll-451.5-1.044×acorrelation452.0 +11.277×ahomogeneiy1352.0+0.014×ahis50mean0 +0.00300×ahis25mean2.5+1.4890×pskewness0-0.19800×pkurtosis0+220.6700×penergy00-0.4230×pkurtosiS1.0+5.7640×phOmOgeneiy.-901.0-13.257×phOmOgeneiy01.5+6.235×pentrOpy451.5 +0.0280×phiS10mean1.50.0590×pkurtosis2.0-3.567×phomogeneiy452.0+0.005×phis10mean2.0 -0.1370 x pcontrast902.5-1.0370 x phomogeneiy02.5 +0.006 x phis25mean2.5Performance of the imaging biomarkerThere was significant difference between the LM-positive and LM-negative group in the imaging biomarker, either based on the unenhanced CT image, arterial phase image or portal venous phase image. Besides, the hazard scores were higher for the LM-positive group.ROC analysis showed that the AUC of the imaging biomarker concstructed based on the portal venous phase image （AUC=0.814） for the detection of synchronous colorectal liver metastases was higher than that of the unenhanced CT image （AUC=0.754） or the arterial phase CT image （AUC=0.799）. The sensitivity and specificity of the imaging biomarker based on portal venous phase was 76.3% and 77.5%, respectively. Moreover, the combined imaging biomarker showed further improved detection performance （AUC=0.912, sensitivity=75.0%; specificity=92.5%）.5. Validation of the imaging biomarkerStratified analyses showed that there was significant different in the hazard scores calculated based on all of the four imaging biomarkers between the LM-positive group and the LM-negative group with each subgroup stratified by clinical risk factors （p<0.05）。Multivarible analysis showed that each of the imaging biomarker was an independent risk factor for the detection of synchronous colorectal liver metastases, along with the CEA level.Each of the imaging biomarkers and CEA level was independent risk factor in the constructed detection models for the detection of synchronous colorectal liver metastases. Among the four detection models, the model integrating the combined imaging biomarker yielded the highest AUC (AUC=0.926. The discrimination performance of the above model was better compared to that of the single combined imaging biomarker alone （AUC=0.912）.Besides, the model integrating the portal venous phase CT image-based imaging biomarker demonstrated better performance （AUC=0.845） compared to that of the unenhanced CT image/arterial phase CT image-based imaging biomarker.Conclusions:CT image-based imaging biomarker could be applied to detect synchronous colorectal liver metastases as an independent predictor. Among the four imaging biomarkers, the portal venous phase CT image-based imaging biomarker demonstrates better performance compared to that of the unenhanced CT image or arterial phase CT image-based imaging biomarker, while the combined imaging biomarker showed the best prediction performance. Moreover, building a prediction model integrating the imaging biomarker and CEA level could further improve the prediction performance. |
PDF Full Text Request