In this study, we successfully exploited and validated a combined model for predicting the pretreatment expression of c-Met and determined that among 20 HCC patients taking sorafenib, those predicted to be positive for c-Met expression had worse survival than those predicted to be negative for c-Met expression. The preliminary results showed that the infusion radiomics signature derived from AP, VP and DP achieved better efficacy than that from each independent phase in both the training and validation cohorts. Moreover, the combined model that incorporated HBV-DNA, GGT-II and the fusion radiomics signature had the best performance in predicting pretreatment c-Met expression, attaining the greatest AUC of 0.878 in the training cohort and an AUC of 0.851 in the test cohort. In addition, to detect c-Met-positive HCC preoperatively, we created a predictive nomogram conglomerating the radiomics signature and clinical risk factors by a straightforward approach, and total points over 80 were considered high risk of c-Met positivity among 20 HCC patients who had taken sorafenib.
The power and potential of enhanced CT images, which are well known for their noninvasiveness and convenience, are increasingly recognized in the field of HCC. Using invisible image features to predict critical molecules for poor prognosis was the most common method in radiomics [15, 18]. In this study, we set up image acquisition parameters such as voxel intensity discretization and voxel size resampling to reduce the differences in image specifications. Three-dimensional (3D) primary tumour volume as well as full size and shape were utilized to predict c-Met-positive HCC, which may provide more precise tumour heterogeneity information [27, 28]. Moreover, we established five image-based prediction models (AP, VP, DP, infusion radiomics and combined models) for c-Met positivity in HCC. It is worth noting that four features from AP showed relatively higher discrimination ability than the other two phases. AP may reflect more characteristic changes in arteries than VP and DP, so we can find more useful information related to AP, which suggests great value for c-Met prediction in HCC. According to recent radiomics studies, wavelet-based features and grey-level textures can expose special patterns hidden in invisible features. This makes quantification of the spatial tumour heterogeneity at multiple scales possible [29, 30]. We discovered that three features achieved high weights in the radiomic signature from the mass of data. The wavelet transformation, GLRLM and GLDM yielded significant information reflecting macroscopic heterogeneities of c-Met-positive HCCs by fusing images and quantifying the distribution of run lengths and grey levels. In the radiomics models of AP, VP and DP, these texture patterns drastically improved prediction and were in concordance with the biological characteristics of c-Met-positive HCC.
Our findings indicated that HBV-DNA was closely associated with c-Met expression in clinical models. In exciting studies, HBV infection promoted the occurrence and development of HCC [31]. It was reasonable to hypothesize that HBV-DNA positivity (> 103 copies/ml) correlated with c-Met positivity. Moreover, GGT-Ⅱ was determined to be a routine analysis for the detection of liver damage in HCC, and even high-normal GGT-Ⅱ levels are a risk factor for mortality [32]. The higher the extent of GGT-Ⅱ positivity, the more severe the hepatocellular impairment, which leads to tumour invasion and metastasis. Therefore, preoperative GGT-Ⅱ may be positively associated with c-Met expression. In the present study, HBV-DNA and GGT-II were used as predictive markers in the training set of a clinical model and achieved a satisfactory AUC of 0.805.
The preoperative combined model including a radiomics signature and clinical risk indicators yielded the best discriminatory ability; its prediction accuracy was higher than that of the clinical model and independent radiomics models. Furthermore, for clinical application, we proposed a nomogram to calculate the likelihood of c-Met positivity which assisted surgeons in promoting personalized therapeutic regimens in HCC patients.
In this study, we collected the clinical and radiomic information of an additional 20 HCC patients taking sorafenib and predicted the outcome of c-Met based on the combined model. We calculated the scores of two clinical features (the levels of GGT-Ⅱ and HBV-DNA) and fourteen radiomics features based on the nomogram to identify a high risk of c-Met positivity. The results showed that patients with predicted c-Met-negative HCC had longer OS compared with those with predicted c-Met-positive HCC. Llovet et al. [33] asserted that the effect of sorafenib treatment in HCC patients with a high baseline level of HGF was unsatisfactory in terms of OS and time to progression (TTP), indicating that HGF/c-Met is one of the independent prognostic factors for the risk of sorafenib resistance in HCC. By using the preoperative combined model, we predicted c-Met expression and filtered out advanced HCC patients who may have acquired drug resistance caused by overexpression of c-Met after taking sorafenib.
Sorafenib resistance is a hot topic in advanced HCC, and the relationship between drug resistance and radiomics is rarely discussed in existing studies on HCC. In our study, we successfully established a multivariate model based on radiomics to predict sorafenib resistance in advanced HCC by c-Met expression. Specifically, c-Met-positive advanced HCC patients were more likely to have sorafenib resistance than c-Met-negative patients, which led to shortened survival. If selected patients have a high risk of c-Met positivity, taking a selective c-Met inhibitor as an anti-sorafenib-resistant drug will be a good option to prolong survival, but further research is necessary to confirm this hypothesis. To differentiate HCC patients with potential risk of sorafenib resistance, we proposed a nomogram for clinical application. The nomogram could serve as a reference for effective systemic treatment in advanced HCC. Therefore, we believe that tumour heterogeneity hidden in the radiomic signature should partly reflect the real world.
It is worth noting that tissue microarrays were applied to collect pathological and peritumoral tissues from each HCC patient. This technique has several advantages: it can not only define the relationship of putative biomarker expression with clinicopathological features but also save time, costs, and tissue samples [34]. In this study, tissue microarrays are built to enable rapid biomarker validation in heterogeneity associated with radiomics data. Moreover, in the algorithm, we adopted efficient linear regression for large radiomic datasets and achieved great specificity, sensitivity and recall.
Our study has several limitations. First, this study was a retrospective analysis with inherent bias, although tumour imaging data were carefully recorded. Second, the size of the sample was small, and the reliability of radiomics for general applications remains an issue. Third, it was a single-centre study that lacked external validation, and further validation with multi-institutional CT protocols is needed to evaluate generalizability. Finally, further research on the use of c-Met inhibitors in c-Met-positive patients was not included, which might provide valuable new evidence about the relationship between radiomics and sorafenib resistance.
In conclusion, radiomic features obtained from AP, VP and DP images are potential biomarkers for predicting c-Met expression in HCC. Our radiomics-based model has the potential to predict c-Met expression, which provides promising assistance in screening for sorafenib-resistance patients with advanced HCC. The predictive nomogram integrating enhanced CT radiomics characteristics, HBV-DNA, and GGT-Ⅱ achieved satisfactory performance in terms of individualized risk assessment based on c-Met expression and in predicting sorafenib resistance.