Based on Gd-EOB-DTPA MRI and a large number of clinical data, this study established a variety of radiomic models for preoperative MVI status prediction in HCC patients. The results showed that AFP > 400ng/mL, maximum diameter > 5cm, extrahepatic growth pattern, intratumoral bleeding, periatumoral enhancement in arterial stage, higher RVI and R-score were independent risk factors for MVI+, and R-score had the greatest influence weight. Based on multi-sequence MRI images of Gd-EOB-DTPA (including seven sequences of T2WI, DWI, ADC, AP, PVP, TP and HBP), we used different algorithms to construct five machine learning models: RF, LR, rbf-SVM, XGB and linear SVM..
In many studies, radiomics technology is considered to be an important technology for the diagnosis and classification of various diseases, but whether it can actually be applied clinically still needs extensive research(13, 20, 24, 27). In the past, MVI radiomics analysis mostly used dynamic contrast-enhanced CT and contrast-enhanced ultrasound as the original images for feature extraction(24, 28–30). The original image used in this study was DCE-MRI. According to the National Comprehensive Cancer Network (NCCN), various evaluation indexes including sensitivity, specificity and accuracy of DCE-MRI in diagnosing liver cancer are higher than those of dynamic CT and other imaging examinations(31, 32). In addition, MRI can also capture the functional and anatomical information of tumors and has the characteristics of multiple sequences and multiple data postprocessing methods, which can theoretically provide more information than CT.
In previous MVI radiomics studies, only one or two algorithms were used to build multifactor models(35). Xu et al. used multiple logistic regression models, Chong et al. used logistic regression and random forest models, and Jiang et al. used an XGBoost model and a three-dimensional convolutional neural network algorithm in deep learning for modeling. Ni et al. found that different dimensionality reduction methods and modeling methods have a great impact on the performance of radiomics models(36).Therefore, linear support vector machine model (linear SVM), support vector machine model with rbf kernel function (RBF-SVM), logistic regression model (LR), random forest model (RF) and XGBoost (XGB) model were used in this study to construct the prediction model. In order to select the most suitable machine learning model for predicting MVI and build radiomics score (R-score).
To find the best machine learning model for building R-score, we used receiver operating characteristic (ROC) curves to evaluate the predictive performance of each model. It has been proved that rbf-SVM model has the best generalization and prediction performance among multiple machine learning models. Therefore, we calculated the SVM score of each patient based on the selected features as R-score. In the follow-up process, we combined all relevant factors to establish a nomogram combined model based on multi-factor logistic regression and compared it with a single model. Compared with the clinicopathological model, apparent imaging model and radiomics model alone, the combined model had better predictive performance, and the AUC values of the training group and the test group were 0.987 (95%CI: 0.973-1.000) and 0.968 (95%CI:0.920-1.000), respectively. It turns out that the multi-sequence model predicts much better than any other single-sequence model, which is in agreement with the findings of Xu et al.(24).
In the diagnosis and prognosis evaluation of diseases, probability estimates obtained by combining the information of multiple predictors observed or measured in individuals rather than the information of a single predictor can usually provide more reliable diagnosis or prognosis evaluation(33, 34). Therefore, to give the model enough comprehensive prediction information, we collected the preoperative laboratory parameters of liver cancer patients and found that AFP was a hazard predictor of MVI. We speculate that this may be due to the release of more AFP as a result of cancer cells being more easily destroyed by the entry of cancer cells into the blood. However, the independent prediction performance of was not very impressive (test cohort AUC = 0.591, 95% CI), which is consistent with the research results of Lee et al. (7).
In addition, the differentiation grade of cancer cells is significantly related to the MVI status, which may be due to the relatively higher invasiveness of cancer cells with a low differentiation degree and their ease of invading tiny blood vessels, thus causing MVI. Contrary to the results of previous studies(24, 25), we found that age and AST were not independent predictors of MVI, and this inconsistency between existing studies may be related to ethnic differences or differences in imaging methods.
We also found that the largest tumor diameter > 5 cm, extrahepatic growth pattern, intratumoral hemorrhage, peritumoral enhancement in the arterial phase, and RVI were also independent predictors of MVI, which is consistent with the study by Ji Hye Min et al. The RVI defined in this study is based on the RVI model proposed by the Sudeep Banerjee team for predicting MVI based on dynamic CT, but our prediction effect was not as good as theirs (accuracy = 0.886, sensitivity = 0.761, specificity = 0.938). The reason may be that dynamic CT is more suitable for judging RVI than MRI. The specific reason needs further experimental analysis and demonstration.
There are still several limitations of this study. First, the test cohort in this study was from the same center as the training cohort, and subsequent studies need to incorporate imaging data from other medical centers to validate the reliability and generalizability of our model. Second, this study is a retrospective study, and there may be potential selection bias of the experimental subjects(37). Third, we did not combine radiomics with genomics as Eran Segal(38) did, which may be a fruitful direction for future research. Finally, due to various conditions, the follow-up data of the patients could not be fully collected, so the actual prognosis of the patients was not included in the analysis in this study, and the model stopped at the point of predicting the MVI status.
In conclusion, we developed a variety of radiomics models based on GD EOB DTPA MRI imaging and various clinical parameters to predict the MVI status of HCC patients. Each model has good accuracy and significant diagnostic value. Our study shows that quantitative and noninvasive radiation analysis may be an effective means to help clinicians judge the MVI status of patients with hepatocellular carcinoma and take appropriate treatment plans.