Preliminary Study On Predicting Colon Cancer Liver Metastasis By Texture Analysis Based On CT Image

Objective:The purpose of this study was to investigate whether CT texture features of primary colon cancer may be an imaging marker for predicting liver metastasis in colon cancer patients.Methods: Retrospective study(n = 138)three subgroups: group A: concurrent liver metastases(n = 49),group B: metachronous liver metastases(n = 41)and group C:follow-up two years without liver metastases(n = 48).Group B included: Group B1:liver metastasis occurred <6 months(n=16).Group B2: The time of liver metastases occurred within 6-12 months after surgery(n=10).Group B3: The time of liver metastases occurred within 12-24 months after surgery(n=15).All patients included a complete preoperative CT enhancement image.Manually 3D sketching ROI images on the ITK-SNAP software.Applying AK software to extract 396 features on the arterial and portal phase images,respectively,and the ABC three groups of arterial phase,portal phase and combined data were compared.The best classification model was chosen to obtain accuracy,sensitivity,specificity,test set ROC curve,test set AUC value,and 95%confidence interval.The R language was used to compare the combined data of B1,B2,and B3,and the area under the ROC curve(AUC)was used to evaluate the diagnostic performance of texture features in predicting liver metastasis.Results: The texture characteristics of the arterial phase and portal phase of the primary colon cancer of patients in group A,B and C were compared in pairs.The results showed that the training set and test set accuracy of the three groups of patients in the arterial phase and portal phase were greater than 70%,the test set AUC value from the lowest0.771 to the highest 0.902.After modeling the combined data of the ABC three groups of arterial phase and portal phase,the AUC values of each group of test sets were greater than 0.8.The three-category comparison of the B1,B2,and B3 patients showed that the AUC values of the B1,B2,and B3 groups were 0.700,0.617,and 0.758.Conclusion: The machine learning model established by the radiographic features of CT-enhanced scan of primary colon cancer has good diagnostic performance for the three groups of simultaneous liver metastasis,metachronous liver metastasis,and non-transfergroup within two years of follow-up,and has certain Prospects can distinguish between patients with early(<6 months)and late(12-24 months)liver metastases.In addition,the diagnostic performance of the combined modeling of arterial phase and portal phase is higher than that of the single phase.In summary,the texture features of primary colon cancer are expected to be imaging markers for predicting liver metastasis in colon cancer patients.