Introduction

Hepatocellular carcinoma (HCC) as the primary form of liver cancer, accounting for 75%–85%, represents the sixth most prevalent malignancy worldwide.1 Radical surgical resection has a potentially curative effect on patients with HCC.2 However, most patients are typically diagnosed in the intermediate-advanced stage of the disease because of occult symptoms.3 Due to the large tumor load, late tumor stage, and poor general condition, surgical resection is not feasible for these patients. Instead, interventional therapy including transcatheter arterial chemoembolization (TACE) or hepatic artery infusion chemotherapy (HAIC), and systemic therapy should be adopted as the main treatment methods.4 5

Lenvatinib, a multitarget small-molecule tyrosine kinase inhibitor (TKI), has become a first-line therapy in HCC due to its controllable adverse effects (AEs) and promising efficacy.6 A previous clinical study demonstrated that treatment with lenvatinib plus pembrolizumab exhibited an objective response rate (ORR) of 46%, suggesting that combined therapy may have greater conversion potential than monotherapy, to achieve radical surgical resection.7 Furthermore, in the case of intermediate-stage patients with HCC, TACE stands as the well-accepted therapeutic approach.8 9 Various randomized controlled trials have affirmed the potential of TACE in enhancing tumor responses for patients with unresectable HCC, thereby yielding favorable outcomes in terms of survival.10 11 Compared with conventional TACE, HAIC achieved a higher conversion resection rate and could obtain longer survival benefits and a higher quality of life.12 Later, further studies demonstrated that compared with lenvatinib monotherapy or combined with TACE, the triple treatment regimen of lenvatinib plus programmed cell death protein 1 (PD-1) inhibitors and interventional (LPI) therapy achieved a higher conversion resection rate (12.7%–50%) and ORR (69.6%–77.4%) in patients with HCC.13–15 Moreover, a prospective multicenter study showed that lenvatinib plus camrelizumab and TACE had favorable conversion therapy efficacy and manageable AEs in unresectable HCC.16 However, there are still some patients who are non-responsive to the LPI therapy and accompanied by non-negligible AEs such as liver function damage, cardiotoxicity, and thrombopenia in clinical practice. There is an urgent need for novel biomarkers to predict the efficacy of the LPI therapy and identify appropriate candidates.

Radiomics can transform image information into structured data by extracting quantitative high-throughput radiomics features from medical images, and then analyzing these data to construct models to guide clinical decision-making.17 18 As a kind of non-invasive biomarker, radiomics were successfully used for the prediction of diagnosis,19 staging,20 prognosis,21 and therapeutic efficacy of HCC.22 Kong et al developed a model combining clinical data and MRI-based radiomics features to predict the treatment effectiveness of TACE in intermediate-advanced stage HCC.23 Currently, in the research on radiomics prediction models of HCC, a number of radiomics models with favorable performance have been established to predict the efficacy of PD-1 inhibitors or TACE or lenvatinib monotherapy in HCC.24–26 However, to our knowledge, no radiomics models have been constructed to predict the efficacy of LPI therapy in HCC.

Here, machine learning radiomics model was successfully developed and validated to accurately predict tumor response of patients with unresectable HCC receiving LPI therapy, resulting in identifying suitable candidates for clinicians.

MethodsData preparation and patient selection

Our study retrospectively collected the clinical data and pretreatment contrast-enhanced CT (CECT) images of 151 patients with unresectable HCC who underwent LPI therapy from March 2019 to November 2022 in Sun Yat-sen Memorial Hospital. Subsequently, the patients who were part of the study were randomly assigned to two different cohorts in a 2:1 ratio. The training cohort comprised 101 patients, while the validation cohort included 50.

The inclusion criteria for this study are presented below: (a) clinical or histopathological diagnosis of HCC; (b) Child-Pugh grade A/B and Eastern Cooperative Oncology Group performance status (ECOG-PS) score 0–1; (c) patients received LPI therapy, and no other anticancer treatments (radiotherapy, etc); (d) patients performed CECT within 1 month prior to LPI therapy and at least one measurable target lesion was available for the tumor response evaluation. The specific criteria for exclusion in our study were (a) a combination of other malignant cancers; (b) invasive HCC with no clear boundaries; (c) lack of crucial clinical data or without post-treatment medical images to evaluate the tumor response; and (d) lost to follow-up.

The patients’ clinical-radiomics information was collected and evaluated from the electronic medical record system in our institution. Clinical characteristics included the participants’ age, sex, levels of blood biochemical indices, liver function indicators, body mass index, ECOG-PS score, Barcelona Clinic Liver Cancer (BCLC) stage, tumor burden indices and current medication regimen. The radiomics data included Digital Imaging and Communications in Medicine (DICOM) format CECT images.

Lenvatinib and PD-1 inhibitors administration

The treatment regimen was determined by the experienced clinician based on the physical condition of patients and in consultation with patients. Participants initially received the treatment of intravenous anti-PD-1 antibody and oral lenvatinib prior to interventional therapy. Oral lenvatinib was administered at a dose of 12 mg daily for patients weighing ≥60 kg or 8 mg daily for patients weighing <60 kg. All participants received intravenous PD-1 inhibitor treatment every 3 weeks (including pembrolizumab 200 mg, sintilimab 200 mg, toripalimab 240 mg, tislelizumab 200 mg, or camrelizumab 200 mg) during lenvatinib oral therapy. Treatment was discontinued when the patients suffered from disease progression or intolerable AEs or successful conversion therapy. The drug dosage was reduced or interrupted when the patient underwent grade 3 or greater treatment-related AEs. The classification of drug-related AEs was based on the criteria of the Common Terminology Criteria for Adverse Events (V.5.0). Simultaneously, patients who were found to be positive for hepatitis B surface antigen (HBsAg) were administered either entecavir or tenofovir alafenamide antiviral therapy.

Interventional therapy implementation

All enrolled patients received either TACE or HAIC during lenvatinib plus PD-1 inhibitors sequential therapy. TACE was performed on demand when there was a significant blood supply around the tumor on CECT or enhanced MRI images. Lenvatinib was discontinued for 3 days pre-TACE and post-TACE. HAIC with a regimen of oxaliplatin, leucovorin, and 5-fluorouracil (FOLFOX) was administered after 1–2 d each cycle of lenvatinib and PD-1 inhibitors (average 4–6 doses).

The procedure of TACE was to puncture the right femoral artery using the Seldinger technique, then insert the vascular sheath, and introduce the catheter into the superior mesenteric artery and celiac trunk for angiography to determine the tumor-feeding arteries. Subsequently, a microcatheter was super-selectively placed into the nutrient arteries of the tumor and slowly injected with an appropriate amount of iodized oil mixed emulsion (lipiodol plus epirubicin or lobaplatin). Ultimately, the contrast agent iodixanol was mixed with gelatin sponge particles and injected into the blood vessel until the blood flow was completely blocked.

HAIC was implemented by super-selective placement of a microcatheter into the tumor-feeding arteries by the same procedure that was described above, with the tip of the catheter left in the nutrient arteries and the tail fixed in the inguinal region. FOLFOX was adopted as the infusion chemotherapy regimen of HAIC every 3 weeks. On the initial day, the dosage of oxaliplatin was 85 mg/m2, leucovorin was 400 mg/m2, and 5-fluorouracil was 400 mg/m2. This was followed by the continuous infusion of 5-fluorouracil at a dose of 2400 mg/m2 for a duration of 46 hours.

Tumor response and follow-up

According to the modified Response Evaluation Criteria in Solid Tumors (mRECIST) criteria, two experienced hepatobiliary surgeons (CS and XL) evaluated the treatment response on the basis of changes in the target lesions on pretreatment and post-treatment imaging examinations (including CECT and enhanced MRI).27 Tumor response was graded as partial response (PR), complete response (CR), progressive disease (PD), and stable disease (SD) based on the alteration in diameter of the target lesions during arterial phase enhancement and the presence or absence of new lesions. Moreover, the focus of this investigation was to predict the treatment response to LPI therapy for patients with HCC. Simultaneously, the participants who showed CR and PR were classified as “responders”, whereas those with PD and SD were categorized as “non-responders”. The ORR refers to the proportion of patients who showed PR or CR to LPI therapy; the DCR is the proportion of patients showing SD, PR, and CR.

The enrolled patients were followed up every 4–12 weeks, and a CECT or enhanced MRI examination was performed to evaluate the treatment response for patients with HCC receiving LPI therapy. The endpoint of follow-up was set for May 22, 2023. Overall survival (OS) refers to the duration from the commencement of treatment until the occurrence of mortality from any cause; progression-free survival (PFS) is defined as the period from the initiation of treatment to disease progression or mortality.

CECT images acquisition, region of interest segmentation, and radiomics features extraction

The pretreatment DICOM format CECT images were acquired from the picture archiving and communication system. All patients underwent CECT scanning (Revolution EVO/ Discovery CT750HD, GE Medical Systems; SOMATOM Force, Siemens Medical Systems) within a month prior to the LPI therapy. The tube voltage of the CT scanner was 120 KV with 200 mAs tube current. The slice thickness of CECT images was 1.25 mm (median 1.25 mm; range 1.0–1.5 mm), and the images were reconstructed on a matrix of 512×512 pixels. Moreover, the CECT images were resampled to 1×1×1 mm image spacing to eliminate the difference in CT images obtained by different CT instruments. The detailed parameters of the CT scanner are displayed in online supplemental table S1.

Subsequently, the three-dimensional (3D) region of interest (ROI) was manually segmented (figure 1A), layer by layer, by two experienced radiologists through 3D slicer software (V.5.2.2, https://www.slicer.org). In case of disagreement, the third senior radiologist was consulted and made the necessary adjustments. Considering that multiple tumor lesions meant a larger tumor burden, for patients with multiple tumor lesions, the two largest lesions with a diameter greater than 3 cm were segmented on arterial phase CECT images as ROI.

Figure 1

Workflow of the clinical-radiomics model development. (A) ROI segmentation, the largest tumor region was segmented layer by layer as the ROI in the arterial phase of contrast-enhanced CT. (B) Radiomics feature extraction, 1223 features were extracted by the radiomics extension module on three-dimensional slicer, including histogram, texture, shape, and filtering features. (C) Radiomics feature selection, 14 radiomics features with the highest correlation were selected through independent sample t-test, Spearman’s correlation test and 10-fold cross-validation LASSO. (D) Radiomics model construction, 14 selected radiomics features were used to establish the radiomics model through the three-layer hidden neural network of MLP. (E) Clinical-radiomics model establishment, the clinical variables with statistical significance in univariate and multivariable analysis were constructed into the clinical model by LR algorithm. Finally, the clinical and radiomics scores were integrated to obtain the clinical-radiomics score for the construction of combined model. BMI, body mass index; ECOG, Eastern Cooperative Oncology Group; HBsAg, hepatitis B surface antigen; ROI, region of interest; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; PT, prothrombin; AFP, alpha-fetoprotein; ALT, alanine aminotransferase; AST, aspartate aminotransferase.

The radiomics extension module on 3D slicer software was used to extract the radiomics features within the ROI,28 which were divided into four parts, namely histogram, texture, filtering, and shape features (figure 1B). We obtained 107 original features initially, namely 18 histogram, 24 Gray Level Co-occurrence Matrix texture, 14 Gray Level Dependence Matrix texture, 16 Gray Level Size Zone Matrix texture (GLSZM), 16 Gray Level Run Length Matrix texture, five Neighborhood Gray-tone Difference Matrix texture features, and 14 shape features. Subsequently, based on the texture and histogram features, 744 wavelet features were calculated from the configured wavelet filtering and 372 LoG (Laplacian of Gaussian) features were transformed from LoG filtering with 1/1.5/2/2.5 kernel sizes. Finally, a total of 1223 features were obtained (online supplemental table S2 and online supplemental figure S1).

Radiomics features selection and clinical-radiomics model construction

To prevent overfitting of the developed radiomics model, the methods used to reduce the feature dimension were depicted as follows. First, the extracted radiomics features were regularized through Z-score normalization before model construction, followed by an independent sample t-test to retain features with p<0.05. Second, Spearman’s correlation was employed to calculate the correlation coefficient among the features. Then, the features exceeding a correlation coefficient threshold of 0.95 were excluded. Finally, the most representative features were selected for the radiomics model establishment through the least absolute shrinkage and selection operator (LASSO) method (figure 1C).

Subsequently, we tested nine different classifiers of supervised machine learning to establish radiomics models based on the selected radiomics features. These classifiers included logistic regression (LR), naive Bayes, k-nearest neighbor (KNN), support vector machine (SVM), random forest (RF), light gradient boosting machine (LightGBM), extra trees, adaptive boosting (AdaBoost), multilayer perceptron (MLP). Among these classifiers, MLP had the highest prediction performance in the validation cohort, thus it was used for further analysis. Meanwhile, the radiomics scores were obtained via the three-layer hidden neural network based on MLP classifiers (figure 1D). Furthermore, we conducted univariate and multivariable analyses of the clinical variables. Those variables that demonstrated statistical disparity (p<0.05) were used to establish the clinical model using the LR algorithm, resulting in the derivation of clinical scores. Finally, the clinical-radiomics model was built by integrating the radiomics and clinical scores through LR (figure 1E). The predictive performance of the three models (clinical, radiomics, and clinical-radiomics) was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC). Moreover, we employed the Delong test to compare the different ROC curves.29 The calibration curve and Hosmer-Lemeshow analysis fitting were plotted to assess the prediction accuracy of the models.30 To further evaluate the clinical application value of our models, we performed the clinical impact curve (CIC) and decision curve analysis (DCA).31

Survival analysis

We also used machine learning-based classifiers to predict prognosis and perform risk ratings for patients with HCC receiving LPI therapy. The best cut-off values developed by the maximum Youden index method and X-tile software were used to categorize the participants into high-risk and low-risk groups. Moreover, Kaplan-Meier (KM) survival curves were used to depict the survival prognosis for OS and PFS. Furthermore, the influence of the radiomics and combined scores on OS and PFS risk stratification was assessed by the HR and concordance index.

Statistical analysis

To develop the prediction models, Python software (V.3.9) was employed. All statistical analyses were conducted by the R program (V.4.3.0, http://www.R-project.org) and SPSS software (V.25.0, IBM). For the analysis of the continuous variables, the independent sample t-test or Mann-Whitney U test was used, while the Pearson χ2 test or Fisher’s exact test was applied for the analysis of the categorical variables. As for survival analysis, the OS and PFS of the patients were evaluated by the KM curve, and the differences in KM curves between the two groups were assessed by the log-rank test. Moreover, the ROC curve and calibration curve were delineated by the R program to evaluate the performance of the prediction models. A significance level of p<0.05 in the two-sided tests was considered statistically significant.

ResultsClinical characteristics and treatment response

The flow chart of this study was demonstrated in (figure 2). In total, 212 patients with unresectable HCC who received LPI therapy were retrospectively reviewed. Subsequently, 61 patients were excluded according to the exclusion criteria, resulting in the enrolment of 151 patients. Among them, 101 patients formed the training cohort and 50 constituted the validation cohort. Moreover, the enrolled patients were treated with lenvatinib, of which 31.8% (48/151) were treated with camrelizumab, 4.6% (7/151) with toripalimab, 27.8% (42/151) with pembrolizumab, 19.9% (30/151) with sintilimab, and 15.9% (24/151) with tislelizumab. As for interventional therapy, 14.6% (22/151) patients underwent TACE, 61.6% (93/151) patients underwent HAIC, and the remaining 23.8% (36/151) patients underwent TACE plus HAIC. The distribution of all aforementioned variables is equally spread across both the training and validation cohorts, with no notable statistical disparities (table 1).

Table 1

Demographic and clinical characteristics of the training cohort and validation cohort

Figure 2

Flow chart of the study design. CECT, CECT, contrast-enhanced CT; ICIs, immune checkpoint inhibitors; uHCC, unresectable hepatocellular carcinoma.

Among the 151 patients receiving LPI therapy, 3 (2.0%) showed CR, 69 (45.7%) showed PR, 33 (21.8%) showed PD, and 46 (30.5%) showed SD. The ORR, DCA, and conversion resection rates were 47.7%, 78.1%, and 23.2%, respectively according to mRECIST criteria (online supplemental table S3). Meanwhile, the distribution of the tumor response showed no statistical distinction between the training and validation sets. Further KM analysis revealed that no considerable disparity was existed in the prognoses of survival between the training and validation cohorts (online supplemental figure S2). Subsequently, we performed the univariate analysis of clinical variables between responders and non-responders. In the training set, the response to LPI therapy was associated with the HBsAg positivity (p=0.015) and Tumor Burden Score (TBS) grade (p=0.037). However, no statistical differences were observed in age, cirrhosis, alpha-fetoprotein, albumin bilirubin grade, vascular invasion, BCLC stage, Child-Pugh stage, and treatment regimen (p>0.05) between the response and non-response groups. Furthermore, further multivariable LR analysis demonstrated that HBsAg (OR 4.049, 95% CI 1.195 to 13.719, p=0.025) and TBS grade (OR 2.294, 95% CI 1.006 to 5.229, p=0.048) were still correlated with the response to LPI therapy in the training cohort (online supplemental table S4). Ultimately, we developed a clinical model by LR algorithm using the aforementioned variables. In the training cohort, the AUC value of the clinical model was 0.669 (95% CI 0.571 to 0.766), whereas in the validation cohort, it was 0.585 (95% CI 0.442 to 0.728). Simultaneously, we also explored the performance of other machine learning classifiers for the clinical model construction, with no statistically significant difference in comparison with LR. The predictive performance of clinical models developed by nine machine learning classifiers is shown in online supplemental table S5.

Selection of radiomics features and performance of radiomics models

We extracted 1223 radiomics features by 3D slicer software in total. Then, the feature dimension reduction methods such as the independent sample t-test, Spearman’s correlation test, and 10-fold cross-validation LASSO were used to prevent model overfitting. Subsequently, we obtained 14 radiomics features (4 LoG features and 10 wavelet features) with the highest correlation used for radiomics model establishment. The selected 14 radiomics features are equally spread across the training and validation cohorts, with significant disparity between the response and non-response groups (figure 3A,B). Moreover, the predicted scores tendency of clinical model and radiomics model tend to be consistent, and the higher the score, the better the therapeutic effect of LPI treatment (figure 4A–C and online supplemental table S6). Meanwhile, the predicted response status of LPI therapy in patients with HCC by radiomics model was highly consistent with the actual tumor response status of patients according to mRECIST criteria in the training and validation sets (figure 4E).

Figure 3

The distribution of 14 selected radiomics features. (A) Heatmap of selected radiomics features distribution between responders and non-responders in the training and validation sets. (B) Violin plot of selected radiomics features distribution between response group and non-response group. *p<0.05, **p<0.01.

Figure 4

Performance and comparison of prediction models in the training and validation cohorts. (A, B) Pod plots of different prediction models among response and non-response groups in the training and validation sets, respectively. The ns represents, p>0.05; *p<0.05, ***p<0.001. (A) represents the training set and (B) represents the validation set. (C, D) Predictive score distribution of the radiomics and clinical-radiomics models between response and non-response groups in the training and validation cohorts. The ***p<0.001. (E, F) Waterfall plot of the predictive score distribution of the radiomics and clinical-radiomics models in the response and non-response groups in the training set. (C, E) represent radiomics score distribution graph; (D, F) represent clinical-radiomics score distribution graph. (G, H) Receiver operating characteristic curve analysis and comparison of different prediction models in the training and validation sets, respectively. (I) Calibration curves of radiomics model and clinical-radiomics model in the training and validation cohorts. (J) Decision curve analysis of the clinical model (green), radiomics model (red) and clinical-radiomics model (blue) in the training cohort. The results show that the net benefits obtained by the different prediction models are greater than two extreme conditions (the treat-all-patients scheme (gray curve) and the treat-none scheme (horizontal black line)). (K, L) CIC shows the actual number of high risks (blue) and the number of high risks predicted by the radiomics model (red) for each risk threshold in the training and validation sets, respectively. The ratio of the blue and red values is the true positive rate. (K) represents the training set and (L) represents the validation set. AUC, area under the curve; CIC, clinical impact curve.

Furthermore, the radiomics model based on the MLP algorithm achieved the highest prediction performance in the validation cohort among nine types of machine learning classifiers, which had an AUC value of 0.900 (95% CI 0.842 to 0.958) in the training cohort and 0.893 (95% CI 0.804 to 0.982) in the validation cohort, respectively. Additionally, the other eight machine learning classifiers also had good prediction accuracy. The AUC values of LR, naive Bayes, KNN, LightGBM, and AdaBoost machine learning algorithms were all >0.7. Meanwhile, SVM, RF, and extra trees achieved AUC>0.8 in the training and validation sets. According to these findings, it was indicated that the efficacy of LPI therapy could be well predicted by machine learning radiomics models. The detailed information on the predictive values of the nine machine learning classifiers is demonstrated in online supplemental table S7 and online supplemental figure S3.

Construction of clinical-radiomics model and clinical application of prediction models

The radiomics model by the MLP algorithm integrated with the clinical model using LR algorithm to obtain a better prediction accuracy clinical-radiomics model, with AUC values of 0.912 (95% CI 0.860 to 0.964) in the training set and 0.892 (95% CI 0.803 to 0.981) in the validation set (figure 4D,F,G,H and table 2). Furthermore, the predictive performance of the clinical-radiomics model and radiomics model was significantly superior to the clinical model in the training and validation cohorts. Nevertheless, the prediction performance exhibited no statistical distinction when comparing the radiomics model to the clinical-radiomics model through the Delong test in the training and validation sets, respectively (online supplemental table S8). The calibration curves showed that the predictive precision of the clinical-radiomics model and radiomics model was highly consistent with the actual observation (figure 4I). As for the clinical application, the DCA demonstrated that the clinical-radiomics model and radiomics model had better net benefits than the clinical model under the risk threshold of about 0–0.9 (figure 4J). The CIC indicated that the radiomics and clinical-radiomics models had high true positive rates and cost-effectiveness under a wide range of risk thresholds in the training and validation sets (figure 4K,L, and online supplemental figure S4).

Table 2

Performance of prediction models for predicting the response to LPI therapy in patients with unresectable HCC in the training and validation cohorts

Radiomics and clinical-radiomics models predicted OS and PFS risk stratification

We successfully classified patients into two risk groups using both the radiomics model and the clinical-radiomics model. In the radiomics model, the HRs of PFS and OS between the high-risk and low-risk groups were 1.913 (95% CI 1.121 to 3.265, p=0.016, cut-off point=0.499, figure 5A) and 4.252 (95% CI 2.051 to 8.816, p=0.001, cut-off point=0.554, figure 5C) in the training cohort, and 2.347 (95% CI 1.095 to 5.031, p=0.012, cut-off point=0.484, figure 5B), 2.592 (95% CI 1.050 to 6.394, p=0.019, cut-off point=0.484, figure 5D) in the validation cohort, respectively (online supplemental tables S9 and S10). Meanwhile, the combined model also had a significant prognostic effect on PFS risk stratification, with HR of 2.160 (95% CI 1.264 to 3.690, p=0.007, cut-off point=0.715, figure 5E) and 2.033 (95% CI 0.965 to 4.285, p=0.039, cut-off point=0.337, figure 5F) in the training and validation sets, respectively. Notably, the validation set did not demonstrate meaningful distinction in terms of OS risk stratification for the combined model (HR 2.181, 95% CI 0.903 to 5.266, p=0.062, cut-off point=0.337, figure 5H), with statistical disparity in the training cohort (HR 3.314, 95% CI 1.600 to 6.865, p=0.003, cut-off point=0.589, figure 5G). Moreover, the 12-month PFS and OS rates of the low-risk group classified by radiomics and clinical-radiomics models were higher than those of the high-risk group, which are detailed in online supplemental tables S11 and S12.

Figure 5

Survival prognosis analysis of the radiomics and clinical-radiomics models for patients with unresectable HCC receiving LPI therapy. (A, B) The Kaplan-Meier curves of PFS between low-risk and high-risk groups in the training and validation sets based on radiomics model. (C, D) The Kaplan-Meier curves of OS between low-risk and high-risk groups in the training and validation sets based on radiomics model. (E, F) The Kaplan-Meier curves of PFS between low-risk and high-risk groups in the training and validation sets based on clinical-radiomics model. (G, H) The Kaplan-Meier curves of OS between low-risk and high-risk groups in the training and validation sets based on clinical-radiomics model. HCC, hepatocellular carcinoma; LPI, lenvatinib plus PD-1 inhibitors and interventional; PFS, progression-free survival; OS, overall survival.

Discussion

The current study explored the potential role of radiomics in predicting the long-term survival and response of patients with HCC receiving LPI therapy. The machine learning radiomics model was developed using pretreatment CECT images to non-invasively identify beneficial candidate populations for LPI therapy and perform risk classification to optimize treatment strategies. Moreover, the performance of the radiomics model was comparable to that of the clinical-radiomics model, which meant that CECT examination alone could non-invasively and effectively predict the patients’ response to LPI therapy, without the need for biopsy or repeated hematological examination. To our knowledge, this is the first study to evaluate the CT-based radiomics model to predict the efficacy of LPI therapy in unresectable HCC.

LPI therapy has demonstrated promising treatment effects in unresectable HCC.13–15 However, due to the tumor heterogeneity and various histological subtypes,32 33 only parts of patients with HCC may benefit from LPI treatment. Meanwhile, the combined regimen of systemic plus interventional therapy also meant more potential probability of treatment-related AEs, which brings unfavorable effects to patients with HCC. Therefore, identifying sensitive individuals who may benefit from LPI therapy is of great significance for optimizing treatment strategies and achieving individualized treatment.

Currently, limited studies have focused on seeking biomarkers to predict the efficacy of LPI therapy in HCC. Lai et al combined TCGA analysis illustrated that high serum levels of CCL28 and betacellulin were associated with the efficacy of the patients receiving lenvatinib combined with toripalimab and HAIC, with longer median OS.34 Moreover, another study showed that the increase of Ig G, Ig λ, and Ig κ in peripheral blood was related to treatment response to TACE plus PD-1 inhibitors and TKI.35 Furthermore, serum procalcitonin levels and inflammatory cytokines VEGF, interleukin (IL)-17, IL-6, and IFN-α were found to be predictive biomarkers for lenvatinib in combination with TACE and PD-1 inhibitors in advanced HCC with large-vessel invasion.36 37 However, the clinical application of the above-reported biomarkers was limited by the biological samples collection, low cost-effectiveness, and the prediction accuracy remained to be further improved. Moreover, relative studies have shown that radiomics can serve as a non-invasive and easily available method to predict treatment response in the past decade.17 18 Whereas, as LPI therapy is an emerging treatment approach, the majority of previous studies have focused on using radiomics to predict the effectiveness of monotherapy or dual therapy in HCC.24 25 38 39

All of the selected radiomics features in our study are LoG (n=4) and wavelet filtering features (n=10) rather than original features, which means that radiomics can identify deeper information that cannot be seen by naked eyes. For example, “GLSZM_Small Area High Gray Level Emphasis” and “GLSZM_Small Area Low Gray Level Emphasis” may imply the overall distribution of high-gray and low-gray level neighborhoods in the tumor ROI, respectively. These radiomics features may be related to tumor heterogeneity and immune microenvironment information,40 41 but further genomic and histopathological data are needed to verify. Previous studies have demonstrated that the treatment efficacy of patients with breast and colorectal cancer is associated with tumor heterogeneity,42 43 which is consistent with our results. Moreover, the clinical-radiomics model had a modestly weaker AUC than the radiomics model in the validation cohort (0.892 vs 0.893, p=0.907), which could potentially be attributed to the inadequate predictive efficacy of clinical scores in the validation cohort.

In terms of clinical applicability, we involved a relatively large population of patients with HCC receiving LPI therapy to construct the radiomics model, with favorable performance. In the training set, the actual response to LPI therapy was successfully predicted in 83.2% (84/101) of patients by the radiomics model. Meanwhile, the radiomics model accurately identified 82.0% (41/50) of the patient actual response to LPI therapy in the validation set. Moreover, the clinical-radiomics model was found to enhance the prediction accuracy, as indicated by the accuracy values of 84.2% (85/101) and 84% (42/50) in the training and validation sets, respectively. The above findings suggested that the radiomics and clinical-radiomics models hold great potential as non-invasive approaches in predicting the response to LPI therapy. Additionally, we assessed the correlation between clinical features and treatment response in the training cohort, demonstrating that HBsAg positive (OR 4.049, 95% CI 1.195 to 13.719, p=0.025) and TBS≥8.83 (OR 2.294, 95% CI 1.006 to 5.229, p=0.048) were correlated with better efficacy.

Notably, survival analysis between responders and non-responders showed that patients who positively responded to LPI therapy had longer OS and PFS than those who did not (online supplemental figure S5), highlighting the importance of accurately predicting the response to LPI therapy. Moreover, the clinical-radiomics model exhibited significant differences in the classification of PFS risk and OS risk in the training set. However, no statistically significant disparity was observed in OS risk classification in the validation cohort, which could potentially be attributed to the small sample size in the validation cohort or insufficient follow-up. In sum, the radiomics model in our study can provide a reasonable basis for the application of LPI therapy in the clinical practical. For an initially diagnosed patient with unresectable HCC, LPI therapy could be considered on condition that the radiomics model indicates the treatment will be effective. However, other kinds of treatment options (eg, bevacizumab plus atezolizumab)44 should be recommended for patients whose unfavorable response was predicted by the radiomics model. As for patients who failed LPI treatment, there are no standard second-line choices. Thus, we advise these patients to accept the internationally recommended second-line treatment (eg, regorafenib),45 which has been approved for the second-line therapy with sorafenib resistance in HCC.

Our study had certain limitations. First, it was a retrospective study conducted in a single center, encompassing a relatively small sample size. The results obtained in the study and the performance of prediction models need to be confirmed by multicenter studies with larger sample sizes. Despite this, our study exhibited the potential of the machine learning radiomics model, as a non-invasive and accurate method, to predict the response to LPI therapy in HCC. Second, 84.1% of the patients in this study had hepatitis B, which may have a certain population selection bias. According to the Chinese subgroup analysis of the REFLECT study,6 lenvatinib may have better therapeutic benefit in the HCC population with hepatitis B viral infection. Therefore, most patients in our study could have an efficacy benefit from LPI therapy. Third, LPI therapy is not approved as first-line treatment for patients with HCC currently. However, recent studies have shown the promising effect and safety of this approach, especially for patients with large tumor loads, as the combination of systemic and interventional therapies may achieve higher ORR and longer survival benefits. In forthcoming investigations, we aim to expand our research scope to explore the prognosis and potential mechanisms of conversion therapy by integrating radiomics with tissue sections obtained from patients who have achieved successful conversion.

In conclusion, a promising radiomics model was successfully constructed and validated using machine learning techniques, demonstrating satisfactory performance in predicting the effect of LPI therapy and OS and PFS risk stratification. It could be used to identify potential candidates for LPI therapy and guide clinical decision-making.