Training cohort: Patients with a diagnosis of HCC who were treated by TACE at the Southern Medical University Nanfang Hospital from March 2009 to March 2016 were included in this study. The inclusion criteria of this retrospective study were as follows: (1) Patients underwent more than two TACE sessions during their therapeutic course (the first two sessions were performed within 3 mo. Then, the rest of the TACE procedures were performed “on-demand”); (2) HCC was diagnosed according to the European Association for the Study of the Liver criteria or histological examination; (3) TACE was performed as monotherapy; (4) patients were ≥ 18 years old and had an ECOG performance status score of ≤ 2 at the time of the first TACE session; (5) patients had HCC at BCLC stage A or B (TACE was chosen by the patient voluntarily after the treatment options had been presented) and well-preserved liver function (Child-Pugh class A or B); and (6) preoperative CT imaging was performed within 0-7 d before the first session of TACE. The exclusion criteria were as follows: (1) Patients had BCLC stage C or Child-Pugh class C disease; (2) Patients had another percutaneous treatment combined with TACE (including radiofrequency/microwave ablation, etc.); and (3) The raw CT data could not be processed (Figure 1). Finally, 137 patients (median age, 54.3 years; range, 19-79 years) were included in the training cohort.

Open in New Tab   Full Size Figure   Download Figure

Figure 1 Flowchart for inclusion and exclusion of patients within the training and validation cohort. TACE: Transarterial chemoembolization; ECOG: Eastern Cooperative Oncology Group; BCLC: Barcelona Clinic Liver Cancer; CT: Computed tomography; HCC: Hepatocellular carcinoma.

Validation cohort: An independent external validation cohort comprised patients who underwent TACE procedures at the Affiliated Hospital of Chengdu University. Consecutive patients who underwent TACE between March 2009 and March 2016 were identified from the local electronic patient recording system. The selection and exclusion criteria for the validation cohort were the same as those for the training cohort. Overall, 81 patients (median age, 57.1 years; range, 21-78 years) were included.

TACE procedure: All TACE procedures were conducted under local anaesthesia via a traditional femoral artery approach. Our method of TACE was consistent with the practical standard method of TACE in Asian countries, which was previously reported[20]. After the first two TACE sessions, both institutions used treatment “on demand”, and TACE was scheduled upon the detection of viable tumours (local or distant intra-/extra-hepatic recurrences) in patients with Child-Pugh A/B, which can still be treated by TACE.

Primary outcome: Overall survival (OS) was defined as the period from 1 d before the first TACE procedure to death or the last follow-up (March 2020). Survival data were obtained by one researcher who was blinded to the study purpose. There was no loss to follow-up in the present study.

TACE refractoriness: In the present study, TACE refractoriness was defined according to the JSH and the Liver Cancer Study Group of Japan (JSH-LCSGJ) criteria[21] as follows: Two or more consecutive insufficient responses of the treated tumour (viable lesion > 50%) or an increase in tumour number after changing the chemotherapeutic agents and/or reanalysis of the feeding artery seen on the response evaluation CT/MRI at 1-3 mo after having adequately performed selective TACE; the appearance of vascular invasion; the appearance of extrahepatic spread; or a continuous elevation in tumour markers immediately after TACE even though a slight transient decrease was observed.

Candidate clinical variables: Based on the ART and ABCR scores, we collected as much clinical data as possible. Before the first TACE session, the presence or absence of a tumour capsule was evaluated according to Liver Imaging Reporting and Data System version 2018 (LI-RADS v2018) ancillary features[22]. Irregular tumour margins were defined as non-smooth margins, presenting as the multinodular confluent type, and nodular type with extranodular extension and/or with an infiltrative appearance[23]. Those radiologic imaging evaluations were assessed by two senior radiologists (more than 10 years in interpreting abdominal imaging) with consensus.

Radiomics feature selection: All patients underwent multidetector CT examination (Siemens Somatom Definition 64; VCT 64; GE Medical Systems). A 1.5 mL/kg body weight bolus of iohexol (Omnipaque, GE Healthcare Co., Ltd.) was injected intravenously at a flow rate of 2.5 mL/s, followed by a 20-mL saline flush. Images of the arterial phase, portal venous phase, and delay phase were obtained at 35 s, 65 s, and 120 s, respectively, after intravenous contrast material injection. The scanning parameters used in this study were as follows: Tube voltage, 120 kVp; detector collimation, 128 mm × 0.625 mm; field of view, 250-400 mm; pixel size, 512 × 512; rotation time, 0.5 s; slice interval, 0 mm; slice thickness, 5 mm; and reconstructed section thicknesses, 1 mm. The processing flow of the radiomics analysis included the following four steps: Tumour segmentation, feature extraction, model building, and model evaluation. Radiomics features were extracted from CT images of the pre-treatment arterial phase. The largest tumours were manually delineated on each transverse image by using ITK-SNAP software (http://www.itksnap.org/pmwiki/pmwiki.php) (Figure 2). Two researchers (with 5 and 8 years of experience in abdominal imaging) who were blinded to the research purpose independently performed tumour delineation and feature extraction with the open-source PyRadiomics package (version 2.2.0). Images were standardized to a voxel size of 1 mm × 1 mm × 1 mm, and voxel intensity values were discretized by using a fixed bin width of 10 HU to reduce image noise and normalize intensities, allowing for a constant intensity resolution across all tumour images[24]. The interobserver reproducibility of the radiomics features was evaluated by the intraclass correlation coefficient (ICC).

Open in New Tab   Full Size Figure   Download Figure

Figure 2 Imaging segmentation. The images were acquired form a 65-year-old man with hepatocellular carcinoma (orange arrows). Serial raw computed tomography scans obtained before transarterial chemoembolization treatment; regions of interest were manually depicted along with the tumor outline on each axial slice and automatically merged into a volume of interest (green area).

The image features and clinical data were evaluated by applying the Student’s t-test, the chi-squared test, or the Mann-Whitney U-test, where appropriate. Missing data were considered at random and handled using multiple imputations by the iterative Markov chain Monte Carlo method (10 iterations)[25]. All statistical analyses were conducted by using R software (http://www.Rproject.org). A two-tailed P < 0.05 was regarded as statistically significant.

Step 1 - construct radiomics score: Radiomics features that had ICCs greater than 0.80 were used for subsequent analysis. We used the least absolute shrinkage and selection operator (LASSO) Cox regression method to determine radiomics features that predicted OS. The radiomics score (Rad-score) cut-off value was determined with the “surv_cutpoint” function of the “survminer” R package.

Step 2 - construct the clinical model/CT-based radiomics nomogram: Univariate and multivariate Cox proportional hazards analyses were used to identify clinical factors related to patient survival. The B regression coefficients of the Cox regression model were multiplied by 2 and rounded to obtain the risk score calculation. A CT-based radiomics nomogram was constructed by incorporating the clinical risk factors and radiomics signature by the use of a multivariable Cox regression model. The radiomics model used the median value of the total score to stratify patients into high- and low-risk groups, and ultimately, a nomogram was built.

Step 3 - model evaluation: The OS of different groups was estimated by using the Kaplan-Meier method and compared by the log-rank test. The performance of each model was evaluated by the concordance index (C-index). The C-index is the area under the curve for continuous time-to-event survival data used to measure the discrimination of a prognostic model. The calibration curve was established, and the Hosmer-Lemeshow test was used to analyse the prognostic performance of the CT-based radiomics nomogram in both cohorts. Decision curve analysis (DCA) was conducted to determine the clinical usefulness of the nomogram in the validation cohort.

Between March 2009 and March 2016, a total of 461 HCC patients underwent TACE as first-line therapy. A total of 137 patients were included in the training cohort, and 81 were included in the validation cohort. The median duration of follow-up was 61.3 mo (interquartile range, 25.5-69.3 mo) for the training cohort and 67.1 mo (interquartile range, 32.4-71.3 mo) for the validation cohort. The median number of TACE sessions was 4 (range, 3-7) in both cohorts. The median survival time for the training cohort was 21.6 mo (95% confidence interval [CI]: 13.6-28.1), while that of the validation cohort was 19.7 mo (95%CI: 12.3-25.2) (training vs validation, P = 0.324).

In the training cohort, 67% of patients were male, and most (80.0%) patients tested positive for HBsAg. The majority of patients were graded as having Child-Pugh A (61%) and BCLC stage B (85%). The proportion of patients with a tumour size of 5 cm or greater was 34%. The included variables were not significantly different between the two cohorts (Table 1).

A total of 869 radiomics features were extracted. Sixty-five radiomics features remained in the arterial phase after the reproducibility test (Figure 3A). All the selected features were entered into the LASSO Cox regression analysis. Eight features were selected to build the radiomics signature using the LASSO regression model (Figure 3B): Two volumetric texture features and six wavelet texture features, respectively. The CT-based radiomics signature can be calculated with the following formula: Radiomics score = 0.174 + 0.432 × Shape_Maximum 3D Diameter - 0.432 × Shape_Elongation + 0.017 × wavelet-HHH_GLSZM_Gray Level Variance - 0.263 × wavelet-HHL_GLSZM_Size Zone Non-Uniformity Normalized + 0.046 × wavelet-HLH_GLSZM_Gray Level Non-Uniformity Normalized - 0.367 × wavelet-HLH_GLSZM_Zone Variance + 0.562 × wavelet-HHH_GLRLM_Low Gray Level Zone Emphasis - 0.214 × wavelet- LHL_GLDM_Dependence Non-Uniformity Normalized. The median Rad-score was 0.75 (range, -1.3-1.7). The cut-off value of the Rad-score was 0.1.

Open in New Tab   Full Size Figure   Download Figure

Figure 3 Feature selection. A: Histogram of the inter-class correlation coefficient. The intraclass correlation coefficient (ICC) was used to determine the stability of the features. Features with an ICC < 0.8 were excluded from the analysis. After robustness test, 65 of the initial 869 computed tomography image features in the arterial phase were retained; B: Partial likelihood deviance was drawn vs log (λ) in the arterial phase.

Clinical model: A multivariate model based on Cox analyses revealed four clinical factors affecting OS: BCLC stage, irregular tumour margin, largest tumour size, and tumour number. These factors were combined to create a clinical score (Table 2).

CT-based radiomics nomogram: The results of the univariate analysis showed that Rad-score, BCLC stage, irregular tumour margin, largest tumour size, tumour number, irregular tumour margin, and AFP level had significant effects on OS (Table 3). Multivariate Cox regression analysis showed that the following five factors affected OS: Rad-score, BCLC stage, irregular tumour margin, largest tumour size, and tumour number (Table 4). The median value of the CT-based radiomics nomogram was 3.5 (Figure 4). We used the median value to classify our patients into high- and low-risk groups.

Open in New Tab   Full Size Figure   Download Figure

Figure 4 The computed tomography-based nomogram was built. The computed tomography (CT)-based nomogram was obtained by combining the effective clinical factors and radiomics signature of artery phase contrast-enhanced CT images. We choose the median value (3.5) to classify our patients into high and low-risk groups. BCLC: Barcelona Clinic Liver Cancer.

The CT-based radiomics nomogram stratified the patients in the training cohort into two subgroups with distinct prognoses. Patients with a score ≤ 3.5 (low-risk group) had a median OS of 23.6 mo (95%CI: 16.3-28.7), whereas patients with a score > 3.5 (high-risk group) had a median OS of 12.3 mo (95%CI: 7.9-16.4; P < 0.001). In the validation cohort, the same trend was observed. The Hosmer-Lemeshow test demonstrated that there was no significant difference between the predictive curve (CT-based radiomics nomogram) and the ideal curve in the training and validation cohorts (P = 0.137 and 0.165, respectively) (Figure 5A and B). DCA indicated that the CT-based radiomics nomogram adds more net benefit than the other four models/scoring systems in the validation cohort (Figure 5C). We used the ART and ABCR scores to identify TACE refractoriness in our cohorts. In both the ART and ABCR groups, the median OS duration did not show a significant difference in either the training or validation cohorts (Figure 6). The C-indexes of the CT-based radiomics nomogram were 0.844 (0.762-0.901) and 0.831 (0.742-0.881) in the training and validation cohorts, respectively, and the nomogram significantly better predicted OS than the other models (the ART score, the ABCR score, our newly constructed clinical model, and the radiomics score) (Table 5).

Open in New Tab   Full Size Figure   Download Figure

Figure 5 Calibration curve and clinical utility analyses. A: Calibration curves of the nomogram on the training dataset. The Hosmer-Lemeshow test yielded a P value of 0.137 in the training dataset; B: Calibration curves of the nomogram on the validation dataset. The Hosmer-Lemeshow test yielded a P value of 0.165 in the validation dataset; C: Decision curve analysis for each model in the validation dataset. The Y-axis measures the net benefit, which is calculated by summing the benefits (true-positive findings) and subtracting the harms (false-positive findings). The decision curve showed that if the threshold probability is over 10%, the application of computed tomography-based radiomics nomogram to predict transarterial chemoembolization-refractoriness adds more benefit compared with the other scores/models. ART: Assessment for Retreatment with Transarterial Chemoembolization; ABCR: α-Fetoprotein, Barcelona Clinic Liver Cancer, Child-Pugh, and Response; CT: Computed tomography.

Open in New Tab   Full Size Figure   Download Figure

Figure 6 Kaplan-Meier curves of survival outcomes in different patient groups. A: Kaplan-Meier analysis of overall survival of high-risk (> 3.5) and low-risk (≤ 3.5) patients according to computed tomography (CT)-based radiomics nomogram in the training cohort; B: Kaplan-Meier analysis of overall survival of high-risk (> 3.5) and low-risk (≤ 3.5) patients according to CT-based radiomics nomogram in the validation cohort; C: Kaplan-Meier analysis of overall survival of high-risk (≥ 4) and low-risk (< 4) patients according to α-fetoprotein, Barcelona Clinic Liver Cancer, Child-Pugh, and Response (ABCR) score in the training cohort; D: Kaplan-Meier analysis of overall survival of high-risk (≥ 4) and low-risk (< 4) patients according to ABCR score in the validation cohort; E: Kaplan-Meier analysis of overall survival of high-risk (≥ 2.5) and low-risk (0-1.5) patients according to Assessment for Retreatment with Transarterial Chemoembolization (ART) score in the training cohort; F: Kaplan-Meier analysis of overall survival of high-risk (≥ 2.5) and low-risk (0-1.5) patients according to ART score in the validation cohort.

The point at which transarterial chemoembolization (TACE) should be continued or stopped is currently not addressed by any guideline to our knowledge. Repeated TACE cycles are, however, associated with an increase of related side effects and liver damage, potentially preventing an even greater survival advantage. In the era of personalized oncology, radiomics has allowed digitally encrypted medical images to be transformed into high-throughput quantitative features that provide information on patient prognosis.

In previous studies, patients with high Assessment for Retreatment with Transarterial Chemoembolization (ART) or α-fetoprotein, Barcelona Clinic Liver Cancer, Child-Pugh, and Response score (ABCR) scores tended to have a poor prognosis. Nonetheless, in terms of predictive ability, neither score was reliable enough to allow for clinical decision-making. Although previous studies have shown the prognostic value of computed tomography (CT) radiomic features for different cancer sites, there is scarcity of multi-centre radiomics research on TACE refractoriness.

The purpose of this study was to develop and validate a CT-based radiomics nomogram for the pre-treatment prediction of TACE refractoriness.

Our study consisted of a training dataset (n = 137) and an external validation dataset (n = 81) of patients with clinically/pathologically confirmed hepatocellular carcinoma who underwent repeated TACE from March 2009 to March 2016. The radiomics features were retrospectively extracted from preoperative CT images of the arterial phase. The radiomics signature was built by least absolute shrinkage and selection operator (LASSO) regression. The CT-based radiomics nomogram incorporating clinical risk factors was built by multivariable logistic regression analysis. The performance of the nomogram was assessed with respect to its calibration, discrimination, and clinical usefulness. We used the concordance index to conduct head-to-head comparisons of the radiomics nomogram with the other four models (ART score, ABCR score, CT-based radiomics signature, and clinical model).

Eight features were selected to build the radiomics signature using the LASSO regression model. The CT-based radiomics nomogram included the radiomics score (HR = 3.9, 95% confidence interval: 3.1-8.8, P < 0.001) and four clinical factors and classified patients into high-risk (score > 3.5) and low-risk (score ≤ 3.5) groups with markedly different prognoses (overall survival: 12.3 mo vs 23.6 mo, P < 0.001). The accuracy of the nomogram was considerably higher than that of the other four models (ART score, ABCR score, CT-based radiomics signature, and clinical model). The calibration curve and decision curve analyses verified the usefulness of the CT-based radiomics nomogram for clinical practice.

The CT-based radiomics nomogram is valuable in preoperatively predicting TACE refractoriness, which may aid interventional radiologist in determining the optimal treatment approach.

First, additional information, such as gene sequence data or the molecular pathway, might be necessary for better interpretation of radiomics features, and this issue is left for future research. Second, larger prospective multicenter studies are needed to externally validate our newly constructed model in the future.