Primary liver cancer (PLC) is the sixth most common malignant cancer and the fourth leading cause of cancer-related death worldwide (Bray et al., 2018). A total of 854,000 new cases are diagnosed every year, and almost 50% of PLCs come from China, which places a heavy burden on Chinese health care (Chen et al., 2018). Hepatocellular carcinoma (HCC) accounts for 75–85% of PLC, which cause a heavy burden of death in China. Approximately 70% of patients with HCC are diagnosed with advanced HCC and thus have missed the best opportunity for surgery (European Association for the Study of the Liver, 2018). The 5 years recurrence rate of HCC is as high as 70%. Local interventional strategies such as transarterial arterial chemoembolization (TACE), radiofrequency ablation (RFA) and radioactive particle implantation are widely used in the treatment of advanced HCC (Makary et al., 2020). The combination of TACE and RFA therapy has a synergistic cytotoxic effect on HCC and results in better local tumour control and longer survival than TACE or RFA alone (Peng et al., 2013; Shimose et al., 2019; Chang et al., 2020; Yuan et al., 2021).

TACE pretreatment reduces or eliminates the “heat sink” caused by blood flow during the RFA procedure and clearly shows lesion contours on CT because of lipiodol deposition. These factors may broaden the extent of coagulation necrosis and ultimately result in better outcomes (Iezzi et al., 2016). In addition, the tissue oedema induced by TACE pretreatment enlarges the necrotic area produced by RFA, thereby reducing local recurrence and increasing the safety of the ablation procedure. Use of the TACE procedure after RFA was shown to have a therapeutic effect on the sublethal heating zone of the tumour (Iezzi et al., 2016), which was the main source of local recurrence (Su et al., 2018; Tan et al., 2019). However, the 1-, 3- and 5-years local tumor progression rates reached 9, 40, 55%, respectively (J. H. Kim et al., 2011). What’s more, as a malignant tumor with obvious heterogeneity, the therapeutic effects of TACE combined with RFA were significantly various. Therefore, development of an accurate prediction model for the progression-free survival (PFS) of these patients is urgently needed.

Efforts have been made to select appropriate candidates for TACE and RFA treatment based on the Barcelona Clinic Liver Cancer (BCLC) staging system and on histopathologic grade, AFP and performance status as well as vascular invasion (Makary et al., 2020). In addition, many risk assessment models have been established to predict recurrence after TACE or RFA treatment (Canale et al., 2018; Wang et al., 2019; Han et al., 2020; Kobe et al., 2020). However, candidate patient selection and prognostic assessment models for patients with intermediate and advanced HCC who have been treated with TACE plus RFA have rarely been reported. Thus, an accurate prognostic prediction model is needed for HCC patients after TACE combined with RFA treatment.

Radiomics is a field that involves tumour segmentation, feature extraction and radiomics model building using the high-throughput mining of quantitative image features extracted from images such as magnetic resonance imaging (MRI) and computed tomography (CT) as well as positron emission computed tomography (Lambin et al., 2017; Mayerhoefer et al., 2020). Radiomics approaches have been identified as efficient strategies for predicting the risk of recurrence and survival in HCC patients (Guo et al., 2019; Hui et al., 2018; Kim et al., 2018; Peng et al., 2018). Radiomics analysis has been reported to have specific advantages. It can provide additional, reliable information that radiologists cannot obtain through analysis with the naked eye and can avoid some mistakes made by radiologists who draw conclusions based on their own personal experience. A previous study suggested that a radiomics nomogram served as a non-invasive preoperative prediction method and exhibited favourable predictive accuracy for microvascular invasion in patients with HBV-related HCC (J. Peng et al., 2018). A study by Kim et al. (2018) indicated that the combination of radiomics signature and clinical factors could accurately predict the survival of TACE-treated HCC patients. Our previous study also indicated that the radiomics and clinical indicator-based predictive nomogram can well predict tumor response in intermediate-advanced HCC (Kong et al., 2021).

This study aimed to develop and verify a novel radiomics model including a radiomics signature and clinical risk factors that can be used to predict the PFS of intermediate and advanced HCC patients treated with a combination of TACE and RFA.

Locoregional therapies, including transarterial chemoembolization (TACE) and ablations, are proposed to postpone disease progression (Palmer et al., 2020). Therapy that combines TACE and RFA has been recommended and has proven to be an effective selection for intermediate and advanced HCC (Z. W. Peng et al., 2013). However, the outcomes are not satisfactory for all individuals. Therefore, it is urgent to identify patients who may truly experience a clinical benefit. In this study, we developed a model for predicting PFS based on a radiomics signature and clinical factors. We found that the PFS of patients with a low-risk radiomics signature was higher than that of patients with a high-risk radiomics signature. The C-index of the combined model in the training cohort was higher than that of clinical model [0.722 (95% CI: 0.657–0.786) vs. 0.669 (95% CI: 0.657–0.786), p < 0.001]. The C-index of the combined model in the validation cohort, was higher than that of clinical model [0.821 (95% CI: 0.726–0.915) vs. 0.76 (95% CI: 0.667–0.851), p = 0.004]. The model combining the radiomics signature and clinical factors provided a better evaluation of PFS than the model that considered clinical factors alone.

A radiomics approach was applied to identify more heterogeneous information from images. Radiomics signature extracted from ultrasonic, CT and MRI images have been found to predict recurrence in HCC patients who receive surgery (Akai et al., 2018), liver transplantation (Guo et al., 2019) or TACE (Song et al., 2020) as well as ablation (C. Yuan et al., 2019). In this study, a total of 396 radiomics features extracted by AK software were classified into five clusters: 42 histogram-based features, nine form factor-based features, 11 GLSZM-based features, 180 RLM-based features and 154 GLCM-based features. Histograms and form factor-based features were used to describe the grey distribution and the geometric characteristics of the region of interest. The GLSZM-, RLM- and GLCM-based features reflect the spatial arrangement of the colour or intensity of lesions. Eight potential predictors were selected from the 396 features by shrinking the regression coefficients with the LASSO method. Especially, variance is a parameter of histogram and represents the average of the squared differences from the Mean of image. Cluster Shade_angle0_offset7 is a texture parameter, which groups the similar samples according to their position and, optionally, normal into clusters. Correlation_AllDirection_offset4 is the Correlation that measures the similarity of the grey levels in neighboring pixels, telling how correlated a pixel is to its neighbor over the whole image. Correlation_angle90_offset7 and Haralick Correlation_AllDirection_offset7_SD is the parameters that measure the degree of similarity of the gray level of the image in the row or column direction, representing the local grey level correlation. the greater their value, the greater the correlation. GLCM Entropy_angle90_offset4 shows the amount of information of the image that is needed for the image compression. Inverse Difference Moment_AllDirection_offset1_SD is the local homogeneity. It is high when local gray level is uniform and inverse GLCM is high. Long Run Emphasis_angle90_offset4 is a Parameter of RLM that is generated for each sample image segment as different directions. A Rad-score based on the eight features was calculated for each patient. The cut-off of the Rad-score was calculated, and the patients were classified into high-risk and low-risk groups. Previous studies demonstrated that the Rad-score could predict the recurrence of HCC patients after TACE treatment. We also found that the AUCs for PFS of the radiomics signature of patients in the training and validation cohorts were as high as 0.83 and 0.81, respectively, indicating good sensitivity and specificity for PFS prediction. In addition, the PFS of the low-risk group was higher than that of the high-risk group, further confirming that radiomics signature could be an option for PFS evaluation of HCC patients treated with TACE and RFA.

The BCLC staging system and AFP level have been proven to be efficient in predicting the survival of liver cancer. Consistent with previous studies (Ren et al., 2019; Shimose et al., 2019), we also found that BCLC stage and AFP were related to the PFS of HCC patients treated with TACE and RFA by univariate and multivariate analysis. Other clinical factors such as tumour size, the presence of multifocal lesions, HBV status and neoplasm grade have been recommended for use in evaluating the short- and long-term outcomes of liver cancer. Studies have indicated that tumour size is an independent prognostic factor for recurrence in HCC patients receiving TACE with or without RFA (Sieghart et al., 2015; H. Liu et al., 2016). In contrast, a study by Peng (Z. W. Peng et al., 2013) showed that tumour size was not associated with recurrence-free survival of HCC. In this study, we also found that tumour size was excluded from the multivariate model and was not a prognostic factor for intermediate or advanced HCC. These discrepancies may be due to selection bias.

Studies have demonstrated that TACE shows synergistic effects with RFA in improving tumour response and survival in intermediate or advanced HCC patients (Z. W. Peng et al., 2013; P. Yuan et al., 2021; W. Liu et al., 2020). However, consensus regarding the optimal time interval between TACE and RFA for balancing efficacy and safety remains unclear. A shorter time interval might increase the potential risk of liver dysfunction. Conversely, a longer time interval favours recovery of liver function but may prolong the hospital stay and might decrease the local efficacy of the combination therapy. It is suggested that the use of a time interval between 1 and 30 days could amplify the synergistic effects of the combination therapy (Choe et al., 2014; Iezzi et al., 2016). In this study, we also showed that an interval of 14 days or less was a predictive factor for PFS, consistent with the study by Peng (Z. W. Peng et al., 2013).

A nomogram was used to provide individual predictions; the accuracy of the predictions was verified by calculating the C-index and AUC, and they were analysed based on the calibration curve, the decision curve and the clinical impact curve (Guo et al., 2019; C. Yuan et al., 2019; Song et al., 2020; Zheng et al., 2020). In this study, three factors (BCLC stage, AFP and time interval) with or without a radiomics signature were used to develop clinical and combined prediction models. The AUC of the combined model was higher than that of the clinical model in both the training and the validation cohorts, indicating better performance of the combined model. Thus, a nomogram incorporating the above four factors was constructed. The nomograms in the training and validation cohorts showed favorable consistency, with C-index values of 0.722 (95% CI: 0.657–0.786) and 0.821 (95% CI: 0.726–0.915), respectively. The decision curve suggested that when the threshold probability was approximately 18%, application of the nomogram with the radiomics signature alone for PFS prediction provided a greater benefit than the non-treatment or overall-treatment strategies. The probability of PFS based on a clinical model ranges from 38 to 82%. The net benefits of the radiomics signature alone and the combined nomogram were higher than that of the clinical model alone. Moreover, the outcome was verified by a clinical impact curve.

Admittedly, this study has certain limitations. As a single-centre and retrospective study, there was inevitable selection bias. Almost half of HCC patients in this study were in BCLC stage C, which have great impact on the outcome owing to the fact that the first-line treatment for these patients should be systemic therapy rather than the combination of TACE and RFA. The PFS evaluation did not comprehensively reflected the overall biological behaviours of these patients, especially who with extrahepatic metastasis. A more accurate evaluation for these patients received TACE and RFA treatment should be further investigated in future study. Besides, the sample size was relatively small, and larger datasets and external validation are urgently needed to validate our findings. In the future, large-sample, multicentre cohorts will be examined to verify the prognostic significance of these radiomics signature.

This study provides a radiomics signature-based nomogram of PFS prediction for intermediate and advanced HCC patients treated with TACE and RFA. The radiomics signature was a prognostic risk factor, and the nomogram combined radiomics and clinical factors acts as a new strategy for predicting the PFS of intermediate and advanced HCC treated with TACE plus RFA. These results could be used to facilitate HCC treatment management.