If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Contrast-Enhanced CT Imaging Features Combined with Clinical Factors to Predict the Efficacy and Prognosis for Transarterial Chemoembolization of Hepatocellular Carcinoma
Address correspondence to: H.J., Department of Radiology, The Second Affiliated Hospital of Harbin Medical University, 246 Xuefu Road, Harbin, 150086, China.
Accurate prediction of treatment response to transarterial chemoembolization (TACE) in patients with hepatocellular carcinoma (HCC) is critical for precision treatment. This study aimed to develop a comprehensive model (DLRC) that incorporates contrast-enhanced computed tomography (CECT) images and clinical factors to predict the response to TACE in patients with HCC.
Materials and Methods
A total of 399 patients with intermediate-stage HCC were included in this retrospective study. Deep learning and radiomic signatures were established based on arterial phase CECT images, Correlation analysis and the least absolute shrinkage and selection (LASSO) regression analysis were applied for features selection. The DLRC model incorporating deep learning radiomic signatures and clinical factors was developed using multivariate logistic regression. The area under the receiver operating characteristic curve (AUC), calibration curve and decision curve analysis (DCA) were used to evaluate the performance of the models. Kaplan-Meier survival curves based on the DLRC were plotted to assess overall survival in the follow-up cohort (n = 261).
Results
The DLRC model was developed using 19 quantitative radiomic features, 10 deep learning features, and 3 clinical factors. The AUC of the DLRC model was 0.937 (95% confidence interval [CI], 0.912-0.962) and 0.909 (95% CI, 0.850-0.968) in the training and validation cohorts, respectively, outperforming models established with two signatures or a single signature (p < 0.05). Stratified analysis showed that the DLRC was not statistically different between subgroups (p > 0.05), and the DCA confirmed the greater net clinical benefit. In addition, multivariable cox regression revealed that DLRC model outputs were independent risk factors for the overall survival (hazard ratios: 1.20, 95% CI: 1.03-1.40; p = 0.019).
Conclusion
The DLRC model exhibited a remarkable accuracy in predicting response to TACE, and it can be utilized as a potent tool for precision treatment.
). The Barcelona Liver Cancer System (BCLC) classifies HCC into five stages based on the patient's general condition, tumor status, and liver function (
). Regrettably, the majority of patients with HCC are diagnosed at an intermediate to an advanced stage and miss the opportunity of undergoing radical surgical resection (
). Transarterial chemoembolization (TACE) is the first-line treatment recommended in the guidelines for multifocal, unresectable BCLC stage B HCC without vascular invasion and distant metastases (
). However, due to the heterogeneity of tumors, the efficacy of TACE is uncertain. The accurate prediction of TACE efficacy not only provides treatment plans, but also maximizes the quality of life and prolongs the survival (
). Therefore, a valid and reliable method for early prediction of TACE efficacy is urgently required.
Several studies have shown that preoperative clinical factors, including pre-treatment tumor growth rate, tumor size, and BCLC B sub-classification, were correlated with better treatment response (
Subclassification of BCLC B stage hepatocellular carcinoma and treatment strategies: proposal of modified bolondi's subclassification (Kinki Criteria).
Validation of the six-and-twelve criteria among patients with hepatocellular carcinoma and performance score 1 receiving transarterial chemoembolization.
Gadoxetic acid-enhanced mri features for predicting treatment outcomes of early hepatocellular carcinoma (< 3 cm) after transarterial chemoembolization.
). However, these models do not incorporate quantifiable image information, thereby losing a large number of features regarding tumor spatial heterogeneity.
In recent years, radiomics studies based on contrast enhanced computed tomography (CECT) images have been increasingly used in liver diseases, especially for evaluating the efficacy of local treatment of HCC, and have achieved satisfactory results (
Prediction of therapeutic response of hepatocellular carcinoma to transcatheter arterial chemoembolization based on pretherapeutic dynamic ct and textural findings.
). Nevertheless, for radiomics is concerned, the outlining of region of interest (ROI) relies on manual delineation and the reproducibility of results is poor. Deep learning (DL), in contrast to the manual outlining of features for radiomics, can be used without outlining of contours (
A deep learning model with incorporation of microvascular invasion area as a factor in predicting prognosis of hepatocellular carcinoma after R0 hepatectomy.
Development and validation of a deep learning model for survival prognosis of transcatheter arterial chemoembolization in patients with intermediate-stage hepatocellular carcinoma.
Liu Z, Liu Y, Zhang W, et al. Deep learning for prediction of hepatocellular carcinoma recurrence after resection or liver transplantation: a discovery and validation study.
Predicting the initial treatment response to transarterial chemoembolization in intermediate-stage hepatocellular carcinoma by the integration of radiomics and deep learning.
). However, in these studies, important intraoperative variables, such as tumor staining, embolic materials, and chemotherapeutic agents, were not considered in the training cohort, and therefore, the variables incorporated in the model are still imperfect, resulting in the decreased model performance.
The present study aimed to develop a multimodal integrated model for predicting TACE efficacy based on DL-radiomic and clinical data, in which intraoperative variables were taken for the first time into account. In addition, the ability of deep learning, radiomics and clinical (DLRC) models to predict prognosis of patients was explored as a secondary objective.
MATERIALS AND METHODS
Patients Characteristics
This retrospective study was approved by the Ethics Committee of the Second Affiliated Hospital of Harbin Medical University (Harbin, China), and patient informed consent was waived. Overall, 399 HCC patients who were admitted to the Second Affiliated Hospital of Harbin Medical University were included in the study. All patients were randomly divided into training and validation cohorts in the ratio of 3:1. The inclusion criteria were as follows: patients with HCC of BCLC B as confirmed by imaging or histopathology, and no radiofrequency ablation, chemoradiotherapy, or surgical excision was used during TACE treatment. All patients underwent abdominal CECT before TACE treatment (within 1 month between CT scan images and TACE treatment). In this study patients with poor quality CECT images that interfered with evaluation or analysis were excluded from this study. Similarly, patients with incomplete clinical information were not included in this study, including: age, gender, portal hypertension, diabetes, cirrhosis, hepatitis, hepatitis, alpha fetoprotein (AFP), aspartate aminotransferase (AST), alanine aminotransferase (ALT), albumin (ALB), total bilirubin (TBIL), BCLC B sub-classification, tumor number, tumor size, tumor location, arteriovenous fistula, embolic materials, combined with a chemotherapeutic agent, tumor staining. The patient selection process is shown in Figure S1.
TACE Protocol and Response Assessment
Two interventionalists with more than 10 years of clinical experience performed TACE. Percutaneous puncture of the right femoral artery was carried out using the Seldinger technique. After a successful puncture, a 5F arterial sheath was inserted. Based on the angiographic findings, the tumor vessel was identified, the vessel was fed with a 2.4F/2.7F micro catheter super selective cannula, and embolic and chemotherapeutic drugs were injected. The embolic drugs included iodinated oil emulsion (epirubicin 40 mg + iodinated oil 10 ml completely emulsified), iodinated oil, gelatin sponge microparticles, polyvinyl alcohol particles, blank microspheres and others. Chemotherapeutic drugs were oxaliplatin, fluorouracil, pirarubicin and others. The doses of embolic and chemotherapeutic drugs were calculated based on the tumor diameter (2-3 ml/cm). The procedure was stopped after 5 min, and angiography was performed again to determine whether there was an unmetabolized portion of the tumor. If there was an unmetabolized tumor, the embolization procedure was repeated. Magnetic resonance imaging (MRI)/CECT was performed for patients within 4-6 weeks after surgery to determine whether there was a viable tumor. TACE response was assessed for each patient according to the Revised Response Evaluation Criteria in Solid Tumors (mRECIST) (
Postoperative responses were classified as complete response (CR), partial response (PR), stable response (SD), and progressive response (PD), and outcomes were classified as objective response (CR+PR) and no response (SD+PD). Radiological responses were evaluated independently by two radiologists with more than 5 years of experience and reviewed by a third physician with 8 years of experience when there was disagreement, as shown in Figure S2, and without knowledge of clinical information. Patients received a review every 2-3 months after TACE, including CECT/MRI, AFP level, and liver function. All patients were followed up for overall survival (OS) by telephone or medical records. OS was defined as the interval between the time that patient first received TACE and the time of death from a cause related to this study (June 1, 2022). Censored data were defined as patients who lost follow-up or survived at the last follow-up (June 1, 2022).
CT Imaging and Image Segmentation
CECT images of all patients were acquired before TACE using a 64-row CT scanner (Siemens, Munich, Germany) according to the standard liver scanning protocol. The CT scan parameters were as follows: tube voltage, 120 kV; tube current, 250 mA; scan field of view, (30.0-50.0) cm × (30.0-50.0) cm; matrix size, 512 × 512; window width, 200-450 Hounsfield units (HU); window position, 30-90 HU, layer thickness, 5 mm. A high-pressure injector (OptiVantage. Mallinckrodt Liebel-Frarsheim, Berlin, Germany) was used to inject the contrast agent (Visipaque) via the anterior tibial vein. CECT images at the arterial phase were acquired within 30-35 s. Then the CECT images were resampled and the density was normalized. The obtained images were then uploaded onto the Shenrui research platform (https://keyan.deepwise.com/login/) for target delineation, and the arterial phase was manually and independently delineated by two radiologists with 8 and 10 years of abdominal clinical experience. The ROI included the tumor capsule, avoiding the bile ducts, large blood vessels, and normal liver tissue. The liver tumor capsule is defined as a thin, low-attenuation capsule around nodular HCC on arterial phase images (Supplementary material Appendix E1).
DL Radiomic Features Extraction
Using the arterial phase CECT images, 713 radiomic features were extracted from each ROI by applying Deepwise software (Table S1). These features included first-order statistical features, two-dimensional (2D) features of shape, statistical textural features, and transformed features. DL features were computed by defining the largest representative slice of the tumor ROI as a sliding block of 256 × 256 pixels centered on each pixel on the images. Transfer learning techniques were applied to compensate for the data deficiency, the ResNet18 architecture was used to extract DL features, and a total of 512 features were extracted. DL features were computed by defining the (Supplementary material Appendix E2).
Selection of Features
DL and radiomic features were performed according to the following steps: first, intra/intergroup correlation coefficients (ICCs) were calculated to explore feature recurrence, and then, correlation analysis was performed using R software (the absolute correlation cutoff value was set to 0.9), in order to remove redundant radiomic features. Next, the least absolute shrinkage and selection (LASSO) regression analysis was used to reduce some regression coefficients to zero. The penalty parameter (lambda) was determined using a five-fold cross-validation based on the minimum error criterion.
Establishment of Prediction Models
Clinical factors were collected retrospectively by the PACS system (Supplementary material Appendix E3). In the training dataset, univariate and multivariate logistic regression analyses were performed to screen out clinical factors that were significantly associated with treatment response. Meanwhile, statistically significant clinical variables, DL and radiomics signatures were placed into multivariate logistic regression models (DLRC). Development of a Nomogram, using the Akaike information criterion as a stopping criterion. In addition, the single-signature model (DL signature, radiomic signature) and DL-radiomic model (DLRS) were separately established and their performance was compared (The detailed flow chart is shown in Fig 1).
Figure 1Structure of the DLRC prediction model. (a) Deep learning features were extracted using Resnet-18 along with image histology feature extraction. (b) Single-factor and multi-factor logistic were applied to filter out meaningful features. (c) An integrated model was developed to predict TACE efficacy using clinical information and CT image information. DLRC, deep learning, radiomics and clinical; TACE, transarterial chemoembolization.
Statistical analysis was performed using R 4.0 software. Student's t-test and Kolmogorov-Smirnov test were used to analyze clinical factors. Predictive accuracy, sensitivity, and specificity were measured using receiver operating characteristic (ROC) curve analysis to measure the performance of all established models, and the area under the ROC curve (AUC) was calculated. The effectiveness of the prediction model diagnosis was evaluated by calculating the consistency index (C-index), and the calibration curves were plotted to assess the calibration of all models in the training and validation cohorts, along with the Hosmer-Lemeshow test. The DeLong test was also used to make comparisons between the cohorts. The false discovery rate (FDR) was used to calibrate the results of the Delong test. Net reclassification index (NRI) and integrated discrimination improvement (IDI) were calculated to compare the performance of DLRC and other models. In addition, stratified analysis of the clinical characteristics of all patients was performed. Decision curve analysis (DCA) assessed the clinical utility of the model by quantifying the net benefit at different threshold probabilities. The Akaike information criterion (AIC) was used to assess the fitness of the model. In addition, the relationship between DLRC scores and OS in the follow-up cohort was evaluated using Kaplan-Meier curves. The Youden index was utilized to determine the optimal cut-off value, and patients were divided into two groups: high-risk group and low-risk group. Univariate and multivariate Cox regression analyses were performed to identify potential prognostic factors for OS, and p < 0.05 were considered statistically significant (Supplementary material Appendix E4).
RESULTS
Baseline Characteristics
Baseline characteristics of 399 patients are shown in Table 1. The number of patients with CR+PR was 122 (40.8%) and 43 (43.0%) in the training and validation cohorts, respectively. The percentages of the efficacy of the two groups were balanced. In both cohorts, there were no significant differences between age, gender, cirrhosis, portal hypertension, diabetes mellitus, and hepatitis (p > 0.05). However, Table S2 shows significant differences between the objective response (CR+PR) and no-response (SD+PD) groups in the training cohort in terms of BCLC B subclassification, tumor location, and tumor staining (p < 0.05) (Fig 2a).
Table 1Baseline Characteristics of the Patients in the Training and Validation Cohorts
Figure 2Forest plot of significant factors and the DLRC model. (a) Forest plot of significant factors with a multivariate regression; (b) A nomogram of DLRC model for each factor. DLRC, deep learning, radiomics and clinical.
In the present study, 10 DL and 19 radiomic signatures were selected to establish the prediction models (Table S3 and supplementary material Appendix E5). There was a weak or no correlation between the imaging features of the two selected signatures (Fig S3). The AUC values of the prediction models developed using the DL and radiomic signatures in the validation cohort were 0.857 and 0.782, respectively (Table S4). In addition, the DLRS model combining the two signatures achieved a better performance on all metrics compared to the signature alone.
Model Performance and Validation
DLRC was modeled using multivariate logistic regression to select DL and radiomic features as independent factors and further integrated with clinical factors (Fig 2b and Fig S3). DLRC showed significant differences (p < 0.001) between objective response and no response cohorts (Fig. 3a, b). The performance of DLRC was not affected by age, gender, cirrhosis, AFP, tumor number, or hepatitis. According to the stratified analysis (Fig. 3c-h). Moreover, the AUC of DLRC was superior to the corresponding unimodal characteristics or DLRS in all cohorts (Table S4 and Fig 4 a, b).
Figure 3Box plots showing the correlation of DLRC score between the response and the relationship between treatment response and DLRC score for each subgroup. (a) Training cohort; (b) Validation cohort; (c) Age (d) AFP (e) Cirrhosis (f) Gender (g)Tumor number (h) Hepatitis. ****p < 0.001. DLRC, deep learning, radiomics and clinical.
Figure 4ROC analyses and calibration curve to assess the abilities. (a) ROC analyses of the training cohort; (b) ROC analyses of the validation cohort; (c)Calibration curves of the training cohort;(d)Calibration curves of the validation cohort. ROC, receiver operating characteristic.
After integrating clinical variables, the DLRC model had the lowest AIC value, while the IDI and NRI results showed that the predictive performance of the DLRC model was significantly improved (Table 2). In the validation cohort, there was a significant statistical difference between DLRC and other models (p < 0.05) (Fig S4). The weights of relevant factors in the DLRC model are shown in Table 3. In addition, as shown in (Fig. 4 c, d), the calibration curves showed a satisfactory agreement between the nomogram predictions and the actual probabilities. Subsequently, the DCA in the training cohort showed that the DLRC model has the highest overall net benefit (Fig. 4 e, f).
Table 2Performance of Models
Models
C-Index (95% CI)
Training Cohort
Validation Cohort
AIC
Radiomic
0.613(0.606, 0.621)
0.571(0.548, 0.593)
416.46
Deep learning
0.846(0.841, 0.851)
0.857(0.842, 0.871)
286.72
DLRS
0.864(0.859, 0.868)
0.873(0.859, 0.887)
272.76
DLRC
0.943(0.940, 0.946)
0.916(0.905, 0.927)
188.93
NRI (95%)
p Value
DLRC vs DLRS
0.269 (0.163, 0.376)
0.248 (0.066, 0.430)
< 0.001
IDI (95%)
p Value
DLRC vs DLRS
0.236 (0.185, 0.288)
0.189 (0.108, 0.270)
< 0.001
AIC, Akaike information criterion; C-index, concordance index; DLRS, deep learning and radiomics; DLRC, deep learning, radiomics and clinical; IDI, integrated discrimination improvement; NRI, net reclassification improvement.
Data are presented as concordance index, with 95% CI in parentheses.
A total of 261 patients completed follow-up for OS and were further evaluated for prognostic value of DLRC. The optimal cutoff value for DLRC, as determined by the Youden index, was 0.156. As illustrated in Figure 5a, the Kaplan-Meier curve showed that a DLRC score above 0.165 was significantly associated with a longer OS (log-rank test, p = 0.00038). Notably, after adjusting for clinical variables, DLRC remained as an independent predictor of OS in the multivariate Cox regression analysis (HR: 1.20, 95% CI: 1.03-1.40; p = 0.019) (Table 4). Other variables that were significant in the multivariate Cox regression analysis were tumor location and AFP level (Fig 5b).
Figure 5Kaplan-Meier curves of OS and forest plot for HCC survival. (a) Kaplan−Meier curves of OS using the high-risk and low-risk according to the DLRC score; (b) Forest plot demonstrates the result of multivariate Cox regression analysis of OS. DLRC, deep learning, radiomics and clinical; OS, overall survival.
In this retrospective study, a DLRC model was developed and validated based on DL features, radiomic features, and clinical factors to predict TACE response. Notably, the DLRC model could accurately classify mid-stage HCC patients into two prognostic subgroups, and it was significantly associated with OS. Alternatively, the predicted outcome of the DLRC model was an independent prognostic factor for OS.
The AUC values of DLRS in the validation dataset were higher than radiomic and DL models. Similarly, previous studies on meningioma, glioma, and breast cancer also obtained the same results, in which models that combined DL features and radiomic features outperformed the single deep learning model and radiomic model (
A CT-based deep learning radiomics nomogram for predicting the response to neoadjuvant chemotherapy in patients with locally advanced gastric cancer: a multicenter cohort study.
). In the present study, the correlation analysis of DL features and radiomic features showed a weak or no correlation, as illustrated in the heatmap. This could be attributed to their different performance decisions, as radiomics restricts the prediction performance to a limited number of pre-specified features, while DL captures more complex visual patterns related to the target variable in CT images through the superposition of convolutional layers and nonlinear activation function (
). DL was applied to preprocess the data, and then, DL features were extracted based on the ResNet18 framework to fully exploit the information contained in the hidden layer of the neural network that is relevant to prediction. The final DLRC model was developed using 10 DL features and 19 radiomic features. Among these radiomic features that were mainly used in the textural analysis of features, previous studies have showed that textural features were correlated with survival in patients with HCC after TACE (
Noninvasive imaging evaluation based on computed tomography of the efficacy of initial transarterial chemoembolization to predict outcome in patients with hepatocellular carcinoma.
). This evidence suggests that the combined features enrich the information related to tumor heterogeneity and TACE efficacy.
More importantly, in the DLRC model, the clinically important intraoperative variable, intraoperative tumor staining in TACE, was included, which was accompanied by a significant weighting in the prediction model. Tumor staining, which inhibits progression by cutting off the tumor blood supply, is an important factor in determining the patient's prognosis. The denser the tumor staining, the richer the blood supply, the corresponding increase in contrast perfusion, and the more desirable the embolic effect (
). Meanwhile, tumor location, and BCLC B sub-stage were also found as variables closely correlated with TACE efficacy. Tumor location was associated with a poor treatment response, and multiple branches of the hepatic artery in segments I and IV were frequently anastomosed, which partly improved the blood supply for embolization of the tumor, leading to ineffective TACE and tumor recurrence (
). Previous Studies showed that the efficacy of TACE in patients with different BCLC B sub classification was not identical, probably due to significant differences in liver function across sub-stages (
Prognosis of patients with intermediate-stage hepatocellular carcinomas based on the Child-Pugh score: subclassifying the intermediate stage (Barcelona Clinic Liver Cancer stage B).
). The AIC confirmed that these selected characteristics were complementary rather than redundant. In addition, the NRI and IDI of the combined model with the addition of clinical variables further confirmed the non-over fitting of the DLRC model in the present study. Moreover, tumor size and AFP level were not included in the current study, and tumor size was analyzed mainly because of the linear overlap between BCLC B sub-staging and tumor size, as BCLC B sub-staging takes into account tumor size. Furthermore, a high baseline AFP level could not be associated with initial treatment response, while several studies reported that a decrease in AFP level was associated with treatment response, which was similar to our findings (
), and using multivariate Cox regression analysis, AFP was found as one of the independent influential factors for survival in the present study.
Another important finding is that the DLRC model was not affected by stratification, and the performance of the DLRC model was assessed by OS in addition to treatment response. The results showed that the DLRC model was significantly associated with OS in patients undergoing TACE. DLRC can accurately classify high- and low-risk patients, and in the current study, patients with higher DLRC scores were significantly associated with a longer OS, with a median OS of 21 months. Low-risk patients did not achieve a longer survival compared to high-risk patients, with a median OS of 17 months. In multivariate Cox regression models, DLRC as a separate variable was an independent prognostic factor for OS. Han et al. also compared post-TACE prediction models with hepatoma arterial-embolization prognostic (HAP) scores, emphasizing the importance of radiological response obtained based on mRECIST assessment as an independent risk factor for predicting survival (
Combined sequential use of HAP and ART scores to predict survival outcome and treatment failure following chemoembolization in hepatocellular carcinoma: a multi-center comparative study.
). Early prediction of treatment response to TACE allows for timely alternative treatments, leading to improve survival. Thus, it is feasible to develop treatment plans based on the proposed DLRC model.
The present study has several limitations. First, it was a retrospective study with involvement of a limited number of HCC patients and lack of external validation. However, the experiment was analyzed by five-fold cross-validation to improve the results. In the future, additional prospective data from multiple centers will be collected for analysis. Second, the TACE dose for each patient was not available through the electronic medical record system, and it was impossible to include the TACE dose as a parameter in the model. In the next prospective study, this will be included. Third, the relationship between gene and protein expression and TACE treatment effects was not assessed, indicating the necessity of further study to eliminate the previously-mentioned limitations.
In conclusion, the present study showed that the DLRC model has a promising accuracy in predicting the preoperative treatment outcome of TACE; therefore, the DLRC model has a satisfactory predictive performance and can be used as a potent tool to assist clinicians in the selection of treatment options.
CONCLUSION
The model combining preoperative CECT imaging features and clinical features could well predict the treatment outcome of TACE preoperatively, and the model also successfully divided patients into high-risk and low-risk groups to help clinicians develop reasonable treatment plans.
Authors' contributions
All authors contributed meet the following criteria for authorship. Zhong-Qi Sun: Conception, Study design, Acquisition of data, Analysis and interpretation, Writing - original draft. Zhong-Xing Shi: Conception, Acquisition of data, Formal analysis, Investigation, Methodology. Sheng Zhao: Study design, Formal analysis, Visualization.Yan-jie Xin: TACE procedure implementation, Visualization, Formal analysis. Hao Jiang: Investigation, Visualization, Formal analysis. Jin-ping Li: Investigation, Visualization, Formal analysis. Jia-ping Li: Validation, Substantially revised or critically reviewed the article. Hui-Jie Jiang: Conceptualization, Project administration, Supervision, Validation, Writing - review & substantially revised or critically reviewed the article.
Ethical approval
The study was approved by the Ethics Committee of the Second Affiliated Hospital of Harbin Medical University.
Acknowledgments
The authors gratefully acknowledge the National Key Research and Development Program of China (2019YFC0118100) and National Natural Science Foundation of China (81873910 and 62171167). Project funded by China Postdoctoral Science Foundation (2021MD703827).
Subclassification of BCLC B stage hepatocellular carcinoma and treatment strategies: proposal of modified bolondi's subclassification (Kinki Criteria).
Validation of the six-and-twelve criteria among patients with hepatocellular carcinoma and performance score 1 receiving transarterial chemoembolization.
Gadoxetic acid-enhanced mri features for predicting treatment outcomes of early hepatocellular carcinoma (< 3 cm) after transarterial chemoembolization.
Prediction of therapeutic response of hepatocellular carcinoma to transcatheter arterial chemoembolization based on pretherapeutic dynamic ct and textural findings.
A deep learning model with incorporation of microvascular invasion area as a factor in predicting prognosis of hepatocellular carcinoma after R0 hepatectomy.
Development and validation of a deep learning model for survival prognosis of transcatheter arterial chemoembolization in patients with intermediate-stage hepatocellular carcinoma.
Liu Z, Liu Y, Zhang W, et al. Deep learning for prediction of hepatocellular carcinoma recurrence after resection or liver transplantation: a discovery and validation study.
Predicting the initial treatment response to transarterial chemoembolization in intermediate-stage hepatocellular carcinoma by the integration of radiomics and deep learning.
A CT-based deep learning radiomics nomogram for predicting the response to neoadjuvant chemotherapy in patients with locally advanced gastric cancer: a multicenter cohort study.
Noninvasive imaging evaluation based on computed tomography of the efficacy of initial transarterial chemoembolization to predict outcome in patients with hepatocellular carcinoma.
Prognosis of patients with intermediate-stage hepatocellular carcinomas based on the Child-Pugh score: subclassifying the intermediate stage (Barcelona Clinic Liver Cancer stage B).
Combined sequential use of HAP and ART scores to predict survival outcome and treatment failure following chemoembolization in hepatocellular carcinoma: a multi-center comparative study.
00687-0&id=gr1.jpg){kind=link}
00687-0&id=gr2.jpg){kind=link}
00687-0&id=gr3.jpg){kind=link}
00687-0&id=gr4.jpg){kind=link}
00687-0&id=gr5.jpg){kind=link}