CT-Based Radiomics For Precise Prediction Of Early Recurrence Risk In Hepatocellular Carcinoma After Surgical Resection

PurposeAiming at the clinical challenge that the risk of early recurrence of hepatocellular carcinoma(HCC)is difficult to predict after surgery accurately,this study intends to:First,we train and validate a CT image segmentation model based on the ResUNet network to achieve a fully automatic segmentation of liver tumors and spleen.Secondly,to explore the correlation between tumor radiomics characteristics,spleen radiomics characteristics,and early recurrence of HCC after surgery,and to construct radiomics signatures,respectively.Finally,a risk prediction model for early postoperative recurrence of HCC was built and validated.The incremental value of radiomics imaging biomarkers to the prediction model was explored.Materials and methodsClinicopathological information and preoperative enhanced CT images of 237 HCC patients were retrospectively collected.In the first part of the study,to train and validate the automatic segmentation model based on the ResUNet network,the patients were divided into training set(180 cases)and validation set(57 cases).Using the ITKSNAP software,tumors were manually segmented at the largest cross-sectional level of the liver tumor in the arterial and portal phase images,and the spleen was segmented at the level near the spleen hilum in the portal phase images.The performance of the segmentation model was evaluated using the average Dice coefficient and 95%Hausdorff distance(HD-95).In the second part of the study,patients were divided into training set(130 cases)and validation set(107 cases)by time,tumor radiomics features were extracted from tumor lesions in the arterial phase and portal venous phase images,the spleen radiomics features were extracted from the spleen in the portal venous phase images,and the radiomics features that were highly correlated with recurrence were screened by machine learning methods,the least absolute shrinkage and selection operator Cox regression model was used to construct radiomics signatures,and KaplanMeier survival curves were drawn to evaluate the correlation between the constructed radiomics signatures and early recurrence of HCC after surgery.In the third part,based on the tumor and spleen radiomics signatures built in the second part,a prediction model combining clinicopathological factors,tumor radiomics signature,and spleen radiomics signature was constructed through univariable and multivariable Cox regression model analysis.The performance of the model was evaluated in terms of discrimination,calibration,and clinical value.Nomograms are used to visualize predictive models.ResultsPart 1:In the validation set,the average Dice coefficient of automatic segmentation of spleen is 0.960,HD-95 is 6.766 mm;the average Dice coefficient of automatic segmentation of liver tumors in arterial phase is 0.659,HD-95 is 31.956 mm;the average Dice coefficient of automatic segmentation of liver tumors in portal venous phase is 0.810,HD-95 is 21.210 mm.Part 2:Tumor radiomics signature constructed from four tumor radiomics features and spleen radiomics signature constructed from four spleen radiomics features were associated with early recurrence of HCC after surgery.Patients in the low-risk group identified by tumor and spleen radiomic signatures had significantly higher early recurrence-free survival than those in the high-risk group(all P<0.05).Part 3:The radiomics mixed model combining microvascular invasion,tumor radiomics signature,and spleen radiomics signature showed good discrimination(training and validation C-indexes of 0.780,0.776,respectively)and calibration.Decision curve analysis indicated an excellent overall net benefit for the model.ConclusionCompared with using only clinicopathological and tumor radiomic biomarkers,the inclusion of non-tumor radiomic biomarkers improved the performance of early recurrence prediction models for HCC after surgery.The nomogram combined with clinicopathological and multi-domain radiomics signatures can be used for early recurrence risk stratification in HCC patients after surgery,which is helpful for the formulation of individualized postoperative adjuvant therapy and follow-up strategies for high-risk patients.