Predicting Progression-free Survival in Primary Liver Cancer Patients via Radiomics from DSA during TACE

doi:10.21203/rs.3.rs-4945011/v1

Download PDF

Article

Predicting Progression-free Survival in Primary Liver Cancer Patients via Radiomics from DSA during TACE

https://doi.org/10.21203/rs.3.rs-4945011/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Purpose: This study aimed to establish a radiomics model based on DSA during TACE to predict transcatheter arterial chemoembolization in patients with PLC who have undergone TACE treatment.

Methods: A retrospective cohort of 133 TACE patients split into training (79) and validation (54) sets extracted radiomics features from DSA images, followed by consistency assessment, feature dimension reduction, and computation of the radiomics score (Radscore). Radiomics models, clinical models, and combined radiomics and clinical models were established on the basis of the Radscore and independent clinical risk factors. Goodness-of-fit assessments were performed for all the models, and calibration and decision curves were used to evaluated their calibration ability and clinical utility.

Results: After applying multiple feature reduction methods, 15 radiomics features were ultimately selected to calculate the Radscore. The mPFS in the low-risk group was significantly longer than that in the high-risk group (training: 15.7 vs. 5.5 months, P < 0.001; validation: 9.3 vs. 2.4 months, P = 0.0012). The combined model outperformed both radiomics and clinical models with higher AUCs, better AIC and BIC values, and a higher log-likelihood. Calibration curves and decision curves confirmed its superior predictive accuracy and clinical utility.

Conclusion: The Radscore based on DSA radiomics features can be used to stratify patients into risk groups, and nomograms based on intraoperative DSA radiomics and clinical indicators during TACE constitute a novel strategy to predict PFS in PLC patients receiving TACE treatment.

Health sciences/Oncology/Cancer/Gastrointestinal cancer/Liver cancer

Health sciences/Diseases/Gastrointestinal diseases

primary liver cancer

transcatheter arterial chemoembolization

progression-free survival (PFS)

digital subtraction angiography

radiomics

Primary liver cancer (PLC) is the sixth most common cancer and the third leading cause of cancer deaths globally [1]. Harriet et al. project that by 2040, there will be 1.4 million new cases and 1.3 million deaths from liver cancer, marking a more than 50% increase from 2020 [2]. Due to its subtle onset and challenges in early detection, most patients are diagnosed at advanced stages, often missing the chance for surgery. Currently, transcatheter arterial chemoembolization (TACE) and systemic therapy are the main treatment options [3–5].

TACE is a key treatment for liver cancer and plays a fundamental role in the treatment of patients with intermediate and advanced liver cancer [6]. However, the efficacy of TACE is influenced by various factors such as tumor size, number, blood supply, stage, liver function, and portal vein tumor thrombus, resulting in significant differences in prognosis among different patients [7–9]. Reliable prognostic markers are essential for predicting outcomes, personalizing treatment, and improving patient management.

Radiomics is an emerging noninvasive method that has advantages such as being nontraumatic, easily accessible, and reproducible compared with traditional pathological features and clinical indicators. Through quantitative data analysis of a large number of high-dimensional features, radiomics can identify heterogeneity among tumor lesions, reveal the correlation between imaging features and clinical outcomes, and identify markers that are useful for clinical decision-making [10]. Diagnostic or prognostic models with stable performance can be established via radiomics technology, and their predictive power is significantly better than that of traditional clinical biomarkers, effectively supporting clinical doctors in making clinical decisions [11, 12].

Digital subtraction angiography (DSA) gives surgeons real-time details about tumors, such as their enhancement, shape, and blood supply, helping them assess how well treatments like embolization are working. Previous studies on predicting TACE effectiveness mostly used CT and MRI [7, 13–15], with a focus on the overall tumor or the normal liver tissue surrounding the tumor. However, liver cancer is a highly vascularized solid tumor, and excessive vascularization and arterialization of the tumor blood supply are important characteristics [16]. The efficacy of TACE is closely related to the arterial blood supply of the patient's tumor [3, 17], so Radiomics studies of tumor blood vessels are crucial. During TACE, conventional angiography shows how blood flows to the tumor and its blood vessel patterns, which helps us use DSA images to explore prognostic markers for TACE effectiveness. Right now, using DSA for radiomics to predict liver cancer outcomes is still in the initial exploration stage. Though some studies have used DSA radiomics to predict TACE efficacy and prognosis with some success, they often face limitations, such as relying on subjective judgments or focusing only on overall tumor or blood flow characteristics [18–21]. There is a lack of objective quantitative indicators, the feature extraction methods are relatively simple, and there are subjective differences and uncertainties, resulting in poor reproducibility.

This study aimed to develop a predictive model based on radiomics features from DSA images during TACE to predict progression-free survival (PFS) in patients with PLC who have undergone TACE treatment. The research outcomes are expected to provide more scientific and accurate evidence for clinical practice, helping doctors develop more targeted treatment plans and improving the prognosis and quality of life of patients with PLC.

Study population

Owing to the retrospective nature of the study, the need for informed consent was waived. This study retrospectively collected data from 269 patients with liver space-occupying lesions who underwent interventional treatment at our center between May 2022 and December 2022.

The inclusion criteria were as follows: (1) patients who underwent at least one imaging examination including enhanced CT or MRI, within 2 weeks before TACE treatment; (2) patients with PLC confirmed by pathology or meeting the clinical diagnostic criteria of the American Association for the Study of Liver Diseases (AASLD) 2018 Practice Guidelines for the Management of Hepatocellular Carcinoma; (3) Eastern Cooperative Oncology Group (ECOG) Performance Status (PS) score of 0 or 1; and (4) liver function classified as Child‒Pugh A or B.

The exclusion criteria were as follows: (1) age less than 18 years or greater than 80 years (n = 0); (2) Incomplete DSA imaging data available (n = 64); (3) duplicated study period (n = 28); (4) previous history of liver resection or radiofrequency ablation (RFA) treatment (n = 24); (5) ruptured liver cancer with bleeding or only transcatheter arterial embolization (TAE) treatment (n = 1); (6) co-occurrence of other malignancies or liver lesions as secondary metastatic tumors (n = 6); and (7) loss to follow-up (n = 13).

Finally, 133 patients were selected as study subjects and randomly divided into a training cohort (N = 79) and a validation cohort (N = 54) at a random allocation ratio of 6:4 (Fig. 1).

Procedure of TACE

Each TACE procedure was carried out by an interventional physician with over 10 years of experience. Prior to the treatment, routine DSA examinations of the superior mesenteric and hepatic arteries were conducted to identify tumor-feeding arteries, assess tumor staining, and evaluate the vascular anatomy. Microcatheters were then inserted into the identified feeding arteries, and TACE treatment commenced after confirming catheter placement via angiography. Details of the DSA image acquisition and parameters are available in the supplementary materials.

Follow-up and evaluation

All patients included in the study underwent regular follow-up every 3‒6 weeks after treatment, which was conducted through regular outpatient visits, inpatient reexaminations, and other methods. Patients without disease progression or death were monitored for at least one year, including medical history reviews, clinical assessments, physical exams, lab tests, and imaging. Each follow-up visit was evaluated by two hepatobiliary surgeons with over 10 years of clinical experience. The assessment of patient response to treatment was based on mRECIST 1.1 [22, 23].

During the follow-up period, patients with multiple lesions, large lesions, or suspected residual viable tumors were repeatedly treated with TACE, if no TACE contraindications were found. However, TACE treatment was discontinued in cases of liver function deterioration (Child‒Pugh C), functional status deterioration (ECOG PS > 2), and disease progression assessed by imaging.

Study endpoint

PFS was the primary endpoint of this study and was defined as the time interval from the start of TACE treatment to the first occurrence of PD or death from any cause, whichever occurred first. All patients were followed up until disease progression, death, or December 31, 2023.

Image segmentation and radiomics feature extraction

Two experienced physicians delineated the region of interest (ROI) by reviewing DSA images frame by frame, using preoperative enhanced CT or MRI to select the slice with the clearest tumor vessel visualization and minimal artifacts. Tumor vessels were outlined using the drawing tool in 3D Slicer 5.2.2, based on their distinct intensity in angiographic images. The aim was to accurately capture tumor vessels while minimizing surrounding normal liver blood vessels. For patients with multiple lesions, the ROI was delineated on the tumor with the largest diameter.

To avoid intragroup variation, Physician A reoutlined the ROI for the DSA images of 60 patients via the same method one month later to test intragroup consistency. The outlined ROIs were then used to extract radiomics features via the Slicer radiomics plugin in 3D Slicer 5.2.2.

Feature selection

In this study, all extracted quantitative radiomics features were initially standardized via Z-score normalization. This process transforms data of varying scales into a standardized score for comparative analysis. High-dimensional imaging feature data were subsequently subjected to dimensionality reduction techniques, as detailed in the supplementary materials.

Radiomics signatures

After LASSO regression, the final selected features were multiplied by their corresponding regression coefficients and combined linearly to construct a radiomics signature containing key features and their corresponding regression coefficients. The radiomics score (Radscore) for each patient was calculated.

Risk stratification

To explore the correlation between the Radscore and PFS in the training and validation cohorts, the optimal cutoff value of the Radscore was determined on the basis of the maximum Youden index of the receiver operating characteristic (ROC) curve of the training cohort. Patients in both the training and validation cohorts were then categorized into a high-risk group (Radscore ≥ optimal cutoff value) and a low-risk group (Radscore < optimal cutoff value). Kaplan‒Meier curves were used to analyze PFS for patients in the high- and low-risk groups, and the log-rank test was applied to determine whether there were significant differences between the progression-free survival curves of the high-risk and low-risk groups.

Construction of the predictive model

Using multivariate Cox regression, this study screened for independent risk factors affecting PFS in patients with PLC. In the training cohort, the clinical features of patients were subjected to stepwise forward univariate Cox regression. Predictors with a P-value < 0.1 in the univariate analysis were selected for inclusion in the multivariate Cox regression analysis. Variables with a P-value < 0.05 in the multivariate analysis were identified as independent clinical risk factors associated with PFS.

This study developed three prediction models: a radiomics model using only the Radscore, a clinical model based on clinical risk factors, and a combined model integrating both Radscore and clinical factors to predict PFS for TACE in liver cancer patients.

Model validation

This study employed various methods to evaluate the performance of the established models in terms of discrimination, goodness of fit, accuracy, and other aspects. First, ROC curves based on the three models at 3, 6, and 12 months were plotted, and the area under the receiver operating characteristic curve (AUC) values under the ROC curves of the three models were compared. The Akaike Information Criterion (AIC) score, Bayesian Information Criterion (BIC) score, and likelihood ratio test (LRT) were introduced to evaluate the goodness of fit of the radiomics prediction model, clinical prediction model, and Radscore. Calibration curves were drawn through 1000 resampling iterations via the bootstrap resampling method to assess the agreement between the predicted PFS of the three models and the actual observed PFS. Additionally, decision curves for the three models were plotted, and the predictive performance and clinical utility of different progression-free survival prediction models were evaluated by quantitatively estimating the net benefit of the two studied cohorts. Figure 2 summarizes the overall experimental workflow

Statistical analysis

Continuous variables are presented as the mean (standard deviation) or median (interquartile range). Intergroup differences were analyzed using t-tests or Wilcoxon tests, while categorical variables were examined with chi-square or Fisher's tests. PFS curves were plotted using the Kaplan–Meier method and compared with log-rank tests. Cox regression models analyzed preoperative data, with variables significant in univariate analysis included in multivariate analysis to identify independent PFS prognostic factors. Variables with p < 0.05 were selected to construct the models. The differences among the established models were evaluated via the likelihood ratio test (LRT). A significance level of p < 0.05 was considered statistically significant for all the statistical tests. All analyses were conducted via the "tableone", "psych", "survival", "glmnet", "survivalROC", "stdca.R", "nricens", "survIDINRI", and "rms" packages in R software (version 4.3.0). For all the statistical tests, P < 0.05 was considered statistically significant.

Baseline information and follow-up results

The follow-up results revealed that 48 patients (36.1%) experienced tumor progression during the follow-up period; and 11 patients underwent surgical resection after TACE conversion. Among them, one patient experienced recurrence and disease progression after surgical resection, whereas no signs of tumor recurrence were observed in the remaining 10 patients during continuous dynamic follow-up. By the cutoff date of follow-up, 13 patients (9.8%) had died, and 11 patients (8.3%) achieved CR. As of December 31, 2023, the median progression-free survival (mPFS) of all patients was 13.4 months.

The demographic and clinical baseline characteristics of the patients are presented in Table S1. There were no statistically significant differences in the baseline data between the training and validation cohorts (p > 0.05) (Table S1). During continuous follow-up, 9 (11.4%), 16 (20.3%), 35 (44.3%) and 19 (24.1%) patients achieved CR, PR, SD, PD, or death, respectively, in the training cohort, and the mPFS was 14.7 months. In the validation cohort, 2 (3.7%), 7 (13.0%), 16 (29.6%) and 29 (53.7%) patients achieved CR, PR, SD, PD, or death, respectively, and the mPFS was 7.7 months.

Radiomics feature selection and the radiomics signature

Using the Slicer radiomics plugin of 3D Slicer 5.2.2, quantitative imaging features were extracted from 837 lesions of 133 primary liver cancer patients (Figure S1). These features were subsequently selected through dimensionality reduction and LASSO-Cox regression modeling. Ultimately, 15 stable nonzero coefficient radiomics features were identified as potential prognostic factors for PFS (Table 1, Figure S2). Radiomics features were constructed through linear combination to compute the Radscore for each patient.

Table 1

Non-zero Coefficient Radiomic Features Selected by LASSO Regression
Number	Name of radiomics feature	Coefficient
Feature1	wavelet.HHL_glszm_SmallAreaLowGrayLevelEmphasis	-0.2966
Feature2	wavelet.HLL_ngtdm_Contrast	0.0134
Feature3	wavelet.LLH_glszm_SmallAreaLowGrayLevelEmphasis	0.0178
Feature4	wavelet.LLL_glcm_Imc1	-0.7264
Feature5	wavelet.LHL_glszm_LargeAreaHighGrayLevelEmphasis	0.3612
Feature6	wavelet.HLL_glszm_LargeAreaHighGrayLevelEmphasis	-0.0012
Feature7	wavelet.LHH_glszm_GrayLevelVariance	-0.2129
Feature8	wavelet.LLL_firstorder_Skewness	-0.4710
Feature9	wavelet.HHH_glcm_Correlation	-0.0403
Feature10	wavelet.HHH_glcm_Imc1	-0.9226
Feature11	wavelet.LLH_ngtdm_Contrast	-0.0258
Feature12	wavelet.HLH_firstorder_Median	-0.1348
Feature13	wavelet.LHL_firstorder_Skewness	0.3847
Feature14	wavelet.LHH_firstorder_Skewness	-0.0759
Feature15	wavelet.HLH_glcm_Imc1	-0.1705

The Radscore calculation formula is as follows:

$$\:Radscore={\sum\:}_{n=1}^{15}\left({Feature}_{n}*\:{Coefficient}_{n}\right)$$

Risk stratification on the basis of the Radscore

ROC curves were constructed for the patients in the training cohort on the basis of the Radscore (Fig. 3A). The maximum Youden index of the ROC curve was 0.5458, and the optimal cutoff value of the Radscore was determined to be 2.7949. According to the optimal cutoff value of the Radscore, patients were classified into high-risk (18 patients) and low-risk groups (61 patients) in the training cohort, and 13 high-risk and 41 low-risk patients in the validation cohort. Kaplan‒Meier analysis of PFS between the two groups revealed that the mPFS of the low-risk group was significantly longer than that of the high-risk group, in both the training cohort (15.7 vs. 5.5 months, P < 0.001, Fig. 3B) and the validation cohort (9.3 vs. 2.4 months, P = 0.0012, Fig. 3C).

A comparison of baseline clinical characteristics between the high-risk and low-risk groups on the basis of the Radscore revealed that, in the training cohort, high-risk patients presented higher neutrophil‒lymphocyte ratios (NLRs) (3.9 vs. 2.7, P = 0.009), poorer liver function (P = 0.001), and a greater incidence of vascular invasion (61.1% vs. 29.5%, P = 0.03) (Table S2). In the validation cohort, high-risk patients had significantly poorer liver function (P = 0.027) (Table S3).

Statistical analysis of follow-up outcomes indicated that in the training cohort, there was no statistically significant difference in the objective response rate (ORR) between the high-risk and low-risk groups (P = 0.065), but there was a significant difference in the disease control rate (DCR) (P < 0.001) (Table S2). In the validation cohort, there were no statistically significant differences in the ORR or DCR between the high-risk and low-risk groups (P = 0.569 and 0.108, respectively) (Table S3).

Factors influencing the prognosis of PLC patients after TACE

The results of univariate Cox regression analysis based on clinical baseline data of the training cohort revealed that prior TACE treatment, aspartate aminotransferase (AST), alkaline phosphatase (ALP), NLR, Child‒Pugh liver function status, BCLC stage, tumor number, tumor diameter, portal vein tumor thrombus, and vascular invasion were relevant factors affecting PFS (P < 0.1). Multivariate Cox regression analysis revealed that prior TACE treatment, the NLR, Child‒Pugh score, tumor number, tumor diameter, and vascular invasion were independent risk factors affecting PFS (P < 0.05) (Table 2).

Table 2

COX Regression Analysis of Patient Outcomes in the Training Cohort
Characteristics	Univariate analysis
Characteristics	Hazard ratio	95%CI	p-value	Hazard ratio	95%CI	p-value
Age	0.747	0.312–1.790	0.513
Male sex	1.195	0.358–3.994	0.772
Diabetes	0.818	0.245–2.733	0.744
Hypertension	0.964	0.385–2.413	0.937
Smoke	1.216	0.555–2.666	0.625
Alcohol	0.878	0.394–1.954	0.75
Viral hepatitis	0.647	0.279-1.500	0.31
Prior TACE	2.181	0.994–4.785	0.052	3.425	1.337–8.772	0.01
Number of TACE	1.663	0.694–3.984	0.254
WBC	1.9	0.569–6.350	0.297
PLT	0.527	0.220–1.263	0.151
ALT	1.248	0.566–2.749	0.583
AST	5.155	1.541–17.244	0.008	4.36	0.833–22.827	0.081
ALB	0.801	0.334–1.919	0.618
ALP	2.849	1.189–6.828	0.019	0.75	0.241–2.339	0.621
Serum AFP	0.804	0.452–1.429	0.457
Total bilirubin	1.434	0.428–4.801	0.559
NLR	1.042	1.008–1.078	0.016	1.062	1.014–1.111	0.01
PLR	0.999	0.996–1.003	0.8
PT	0.967	0.721–1.297	0.823
Child-Pugh class	4.743	1.949–11.542	< 0.001	9.163	2.577–32.575	< 0.001
ALBI grade	1.568	0.738–3.333	0.242

Table 2

(Continued)
Characteristics	Univariate analysis
Characteristics	Hazard ratio	95%CI	p-value	Hazard ratio	95%CI	p-value
BCLC stage	2.641	1.410–4.947	0.002	0.747	0.193–2.882	0.672
Number of tumors	2.46	0.982–6.165	0.055	4.274	1.254–14.563	0.02
Tumor diameter	1.796	1.034–3.120	0.038	2.178	1.004–4.722	0.049
Lymph node metastasis	2.438	0.835–7.119	0.103
Extrahepatic metastases	1.147	0.393–3.341	0.802
PVTT	3.323	1.323–8.345	0.011	3.096	0.884–10.836	0.077
Vascular invasion	2.886	1.308–6.370	0.009	6.852	1.325–35.425	0.022
Liver cirrhosis	1.765	0.805–3.870	0.156
Portal hypertension	1.117	0.419–2.977	0.825
Ascites	2.187	0.653–7.323	0.204
Splenauxe	0.762	0.286–2.031	0.587
DEB-TACE	1.439	0.656–3.155	0.364
TT with/without IT	0.902	0.560–1.451	0.669
Follow-up treatment	1.01	0.403–2.531	0.982
Note: WBC: White Blood Cell; HGB: Hemoglobin; PLT: Platelet; ALT: Alanine Aminotransferase; AST: Aspartate Aminotransferase; ALB: Albumin; ALP: Alkaline Phosphatase; AFP: Alpha-Fetoprotein; prior TACE: Previous Transarterial Chemoembolization; NLR: Neutrophil-to-Lymphocyte Ratio; PLR: Platelet-to-Lymphocyte Ratio; PT: Prothrombin Time; ALBI: Albumin-Bilirubin Grade; BCLC: Barcelona Clinic Liver Cancer Staging System; PVTT: Portal Vein Tumor Thrombus; DEB-TACE: Drug-Eluting Bead Transarterial Chemoembolization; TT: Targeted Therapy; IT: Immunotherapy.

Prediction model construction and visualization

Three prediction models were constructed in this study. The radiomics prediction model was built on the basis of the Radscore, and incorporates six independent risk factors affecting the PFS of PLC patients after TACE, including prior TACE treatment, the NLR, Child‒Pugh score, tumor number, tumor diameter, and vascular invasion. A clinical prediction model was constructed by using these independent risk factors. The combined prediction model was developed by integrating the Radscore risk score with independent clinical risk factors affecting PFS in PLC patients after TACE. A nomogram was used to integrate multiple predictive indicators onto the same plane to visualize the three survival prediction models (Fig. 4).

Model evaluation and validation

Figure 5 presents the ROC curves for the radiomics, clinical, and combined radio-clinical prediction models. The analysis reveals that at the three assessed time points in the training cohort, each model achieved AUC values exceeding 0.75, reflecting robust model discrimination. The combined prediction model exhibited superior AUC values compared to both the radiomics and clinical models (Fig. 5), indicating enhanced discriminative performance. These results were corroborated in the validation cohort, where the combined model consistently maintained AUC values above 0.75 at 3, 6, and 12 months, demonstrating its reliable predictive capability across different datasets.

In the training cohort, the combined prediction model exhibited lower AIC and BIC values compared to both the radiomics and clinical prediction models. A log-likelihood ratio test was conducted to compare the three models, with results detailed in Table 3. The log-likelihood value for the combined prediction model (-78.8) was higher than that for the radiomics model (-89.6) and the clinical model (-83.9), with significant differences observed between the combined model and each of the other two models (both P < 0.05). These findings were corroborated in the validation cohort, underscoring that the combined prediction model offers superior fit compared to the radiomics and clinical models.

Table 3

Comparison of the Fitting Effectiveness of Three Models
Cohort	Prediction model	AIC	BIC	LRT
Cohort	Prediction model	AIC	BIC	loglik	X²	p-value
Training cohort	Radiomics nomogram	181.1	182.3	-89.6	21.6	0.001
	Clinical nomogram	179.8	187.1	-83.9	10.3	0.001
	Combined nomogram	171.5	180.1	-78.8
Validation cohort	Radiomics nomogram	247.6	249.2	-122.8	32.1	< 0.001
	Clinical nomogram	238.0	247.3	-113.0	12.4	< 0.001
	Combined nomogram	227.5	238.4	-106.8
Note: AIC: Akaike Information Criterion; BIC: Bayesian Information Criterion; LRT: Likelihood Ratio Test; loglik: Log Likelihood Value; X2: Chi-squared Value.

Figure 6 displays the calibration curves for the three models across both the training and validation cohorts. The results demonstrate that the combined prediction model exhibits a high level of agreement between the estimated values and the actual observed values, with superior calibration compared to both the radiomics and clinical prediction models.

Figure S3 presents the decision curves for the three prediction models. DCA analysis reveals that, within a specific range of threshold probabilities, the combined prediction model demonstrates superior clinical utility in predicting patients' PFS probability compared to both the radiomics and clinical prediction models. At identical threshold probabilities for medical interventions, the combined model offers greater net benefits to patients.

TACE is a standard treatment for patients with intermediate to advanced liver cancer, crucial for disease control, conversion therapy, and bridging to liver transplantation [24]. Given the variable responses and prognoses to TACE among liver cancer patients, predicting efficacy and prognosis post-TACE is essential for evaluating patient response and informing clinical decisions to enhance outcomes. Tumor heterogeneity is a significant prognostic factor in liver cancer [25],and radiomics technology can extract detailed tumor information from medical imaging, highlighting lesion heterogeneity often overlooked by visual inspection [11, 26], This technology can provide insights into tumor biology and microenvironment, serving as a valuable complement to current clinical prediction tools and aiding in treatment decisions.

A recent meta-analysis involving CT or MRI imaging demonstrated that combined imaging-genomic clinical models outperformed standalone imaging or clinical models in predicting survival outcomes [27]. These findings underscore the significant potential of radiomics for predicting the efficacy and prognosis of liver cancer. Moreover, the results of this study surpass those of the meta-analyses, suggesting that intra-TACE DSA radiomics may offer more precise predictions compared to CT or MRI.

DSA provides high-resolution, real-time imaging of tumor morphology and blood supply, which is valuable for both diagnosis and treatment. However, its application in imaging-genomic analyses for liver cancer has been limited due to challenges such as brief image storage durations and restricted accessibility [18, 20]. Several factors contribute to this scarcity of research: the short retention time of DSA images and the difficulty in obtaining DSA images generated by intraoperative CBCT during TACE procedures. Additionally, the field of radiomics research using DSA images from CBCT is still emerging, with no high-throughput studies yet utilizing intra-TACE DSA imaging to analyze liver tumor vascular characteristics for survival predictions. Our study addresses this gap in the literature by exploring the potential of intra-TACE DSA imaging in this context.

Previous research has employed deep learning methods to analyze intra-TACE DSA and pre-TACE CT images for predicting tumor response following the initial TACE treatment for liver cancer. In these studies, the first, which utilized intra-TACE DSA images, reported AUC values of 0.782 and 0.670, while the second, based on pre-TACE CT images, yielded AUC values of 0.981 and 0.972 [18, 28]. The disparity in predictive performance may be attributed to the focus of deep learning models using intra-TACE DSA images on overall tumor features rather than on detailed vascular characteristics. Further research is needed to enhance the deep analysis of DSA vascular features to improve predictions of TACE efficacy.

In this study, patients were categorized into high-risk and low-risk groups based on their radiomics scores. The findings indicated that the median progression-free survival (mPFS) for low-risk patients was significantly greater than that for high-risk patients, with Kaplan‒Meier analysis showing a notable difference in progression-free survival between the two groups. This demonstrates that radiomics scores are valuable for risk stratification and prognosis assessment in liver cancer patients.

A key challenge of the study was the segmentation of vascular images, which required focusing on tumor blood vessels rather than overall tumor features. Additionally, since DSA images obtained during TACE are two-dimensional, the study needed to account for various factors such as gastrointestinal visualization, as well as artifacts caused by heart and diaphragm motion [18]. To address this, we selected the slice with optimal visualization of tumor blood vessels and minimal motion interference around the liver for each patient. We used the gray-scale differences between the blood vessels and surrounding tissues to define intensity ranges, allowing us to delineate the tumor area and extract tumor vascular images directly. However, this method has limitations. For instance, in patients with prior TACE treatments, iodine oil deposition may be present in the DSA images, necessitating that ROI delineation physicians manually remove this interference from the extracted ROI.

This study has the following limitations: First, the study had a small sample size and lacked external validation, potentially limiting the results. Second, its retrospective nature may introduce selection bias, and there was variability in patient characteristics and TACE treatments that was not standardized. Third, most patients had hepatitis B-related liver cancer, which differs from the primary causes of liver cancer observed in Western countries [29]. Future research should focus on larger, multicenter, and prospective studies to address these limitations. Given that liver cancers with different etiologies may display variations in vascular features, it is essential to include a broader range of patients, including those with alcohol-related and hepatitis C-related liver cancer, to enhance the understanding of these differences.

Acknowledgement

We extend our heartfelt gratitude to all patients who participated in this study.

Data Statement

The data sets and materials in this study are available from the corresponding author.

Funding

This study was supported by the Key Research and Development Project of the Science & Technology Department of Sichuan Province [grant numbers: Nos.2022YFS0625].

Ethical statement

This study complied with the Declaration of Helsinki and was approved by the Ethics Committee of the Affiliated Hospital of Southwest Medical University (Ethical Approval Number: KY2024042).

Conflict of Interest

The authors declare no competing interests.

Author Contributions

Guarantor of integrity of the entire study: Yong-fa Liu, Cheng Cui, Bo Li

Study concepts and design: Yong-fa Liu, Bin Luo, Bo Li

Literature research: Ben-jian Gao, Fang-yi Peng, Shuai Hu, Lin-xin Liu

Clinical studies: Yong-fa Liu, Bin Luo, Xue-cheng Bai, Xiao-li Yang

Experimental studies / data analysis: Yong-fa Liu, Cheng Cui, Ben-jian Gao

Statistical analysis: Yong-fa Liu, Cheng Cui, Ben-jian Gao, Ya-ling Li

Manuscript preparation: Yong-fa Liu, Cheng Cui, Ya-ling Li

Manuscript editing: Yong-fa Liu, Cheng Cui, Xiao-li Yang, Bo Li

Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. May 2021;71(3):209-249. doi:10.3322/caac.21660
Rumgay H, Arnold M, Ferlay J, et al. Global burden of primary liver cancer in 2020 and predictions to 2040. J Hepatol. Dec 2022;77(6):1598-1606. doi:10.1016/j.jhep.2022.08.021
Llovet JM, Kelley RK, Villanueva A, et al. Hepatocellular carcinoma. Nat Rev Dis Primers. Jan 21 2021;7(1):6. doi:10.1038/s41572-020-00240-3
Llovet JM, De Baere T, Kulik L, et al. Locoregional therapies in the era of molecular and immune treatments for hepatocellular carcinoma. Nat Rev Gastroenterol Hepatol. May 2021;18(5):293-313. doi:10.1038/s41575-020-00395-0
Benson AB, D'Angelica MI, Abbott DE, et al. Hepatobiliary Cancers, Version 2.2021, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw. May 1 2021;19(5):541-565. doi:10.6004/jnccn.2021.0022
Lu J, Zhao M, Arai Y, et al. Clinical practice of transarterial chemoembolization for hepatocellular carcinoma: consensus statement from an international expert panel of International Society of Multidisciplinary Interventional Oncology (ISMIO). Hepatobiliary Surg Nutr. Oct 2021;10(5):661-671. doi:10.21037/hbsn-21-260
Chen M, Cao J, Hu J, et al. Clinical-Radiomic Analysis for Pretreatment Prediction of Objective Response to First Transarterial Chemoembolization in Hepatocellular Carcinoma. Liver Cancer. Feb 2021;10(1):38-51. doi:10.1159/000512028
Piscaglia F, Ogasawara S. Patient Selection for Transarterial Chemoembolization in Hepatocellular Carcinoma: Importance of Benefit/Risk Assessment. Liver Cancer. Mar 2018;7(1):104-119. doi:10.1159/000485471
Sieghart W, Hucke F, Peck-Radosavljevic M. Transarterial chemoembolization: modalities, indication, and patient selection. J Hepatol. May 2015;62(5):1187-95. doi:10.1016/j.jhep.2015.02.010
Feng Z, Li H, Liu Q, et al. CT Radiomics to Predict Macrotrabecular-Massive Subtype and Immune Status in Hepatocellular Carcinoma. Radiology. Apr 2023;307(1):e221291. doi:10.1148/radiol.221291
Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology. Feb 2016;278(2):563-77. doi:10.1148/radiol.2015151169
Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. Dec 2017;14(12):749-762. doi:10.1038/nrclinonc.2017.141
Kong C, Zhao Z, Chen W, et al. Prediction of tumor response via a pretreatment MRI radiomics-based nomogram in HCC treated with TACE. Eur Radiol. Oct 2021;31(10):7500-7511. doi:10.1007/s00330-021-07910-0
Guo Z, Zhong N, Xu X, et al. Prediction of Hepatocellular Carcinoma Response to Transcatheter Arterial Chemoembolization: A Real-World Study Based on Non-Contrast Computed Tomography Radiomics and General Image Features. J Hepatocell Carcinoma. 2021;8:773-782. doi:10.2147/JHC.S316117
Mao X, Guo Y, Wen F, Liang H, Sun W, Lu Z. Applying arterial enhancement fraction (AEF) texture features to predict the tumor response in hepatocellular carcinoma (HCC) treated with Transarterial chemoembolization (TACE). Cancer Imaging. Aug 12 2021;21(1):49. doi:10.1186/s40644-021-00418-2
Cramer T, Vaupel P. Severe hypoxia is a typical characteristic of human hepatocellular carcinoma: Scientific fact or fallacy? J Hepatol. Apr 2022;76(4):975-980. doi:10.1016/j.jhep.2021.12.028
Forner A, Llovet JM, Bruix J. Hepatocellular carcinoma. Lancet. Mar 31 2012;379(9822):1245-55. doi:10.1016/S0140-6736(11)61347-0
Zhang L, Jiang Y, Jin Z, et al. Real-time automatic prediction of treatment response to transcatheter arterial chemoembolization in patients with hepatocellular carcinoma using deep learning based on digital subtraction angiography videos. Cancer Imaging. May 12 2022;22(1):23. doi:10.1186/s40644-022-00457-3
Wang E, Xia D, Bai W, et al. Tumor Hypervascularity and hand-foot-skin reaction predict better outcomes in combination treatment of TACE and Sorafenib for intermediate hepatocellular carcinoma. BMC Cancer. Apr 30 2019;19(1):409. doi:10.1186/s12885-019-5570-z
Miller JP, Ramaswamy R, Akinwande O. Using principal component analysis for the prediction of tumor response to transarterial chemoembolization. Abdom Radiol (NY). Jul 2019;44(7):2594-2601. doi:10.1007/s00261-019-01982-9
Becker LS, Maschke SK, Dewald CLA, et al. Two-dimensional parametric parenchymal blood flow in transarterial chemoembolisation for hepatocellular carcinoma: perfusion change quantification and tumour response prediction at 3 months post-intervention. Clin Radiol. Feb 2021;76(2):160 e27-160 e33. doi:10.1016/j.crad.2020.09.012
Lencioni R, Llovet JM. Modified RECIST (mRECIST) assessment for hepatocellular carcinoma. Semin Liver Dis. Feb 2010;30(1):52-60. doi:10.1055/s-0030-1247132
Llovet JM, Lencioni R. mRECIST for HCC: Performance and novel refinements. J Hepatol. Feb 2020;72(2):288-306. doi:10.1016/j.jhep.2019.09.026
Minici R, Siciliano MA, Ammendola M, et al. Prognostic Role of Neutrophil-to-Lymphocyte Ratio (NLR), Lymphocyte-to-Monocyte Ratio (LMR), Platelet-to-Lymphocyte Ratio (PLR) and Lymphocyte-to-C Reactive Protein Ratio (LCR) in Patients with Hepatocellular Carcinoma (HCC) undergoing Chemoembolizations (TACE) of the Liver: The Unexplored Corner Linking Tumor Microenvironment, Biomarkers and Interventional Radiology. Cancers (Basel). Dec 30 2022;15(1)doi:10.3390/cancers15010257
Li L, Wang H. Heterogeneity of liver cancer and personalized therapy. Cancer Lett. Sep 1 2016;379(2):191-7. doi:10.1016/j.canlet.2015.07.018
Lubner MG, Smith AD, Sandrasegaran K, Sahani DV, Pickhardt PJ. CT Texture Analysis: Definitions, Applications, Biologic Correlates, and Challenges. Radiographics. Sep-Oct 2017;37(5):1483-1503. doi:10.1148/rg.2017170056
Wang Y, Li M, Zhang Z, Gao M, Zhao L. Application of Radiomics in the Efficacy Evaluation of Transarterial Chemoembolization for Hepatocellular Carcinoma: A Systematic Review and Meta-analysis. Acad Radiol. Jan 2024;31(1):273-285. doi:10.1016/j.acra.2023.08.001
Peng J, Huang J, Huang G, Zhang J. Predicting the Initial Treatment Response to Transarterial Chemoembolization in Intermediate-Stage Hepatocellular Carcinoma by the Integration of Radiomics and Deep Learning. Front Oncol. 2021;11:730282. doi:10.3389/fonc.2021.730282
Kim DY, Ryu HJ, Choi JY, et al. Radiological response predicts survival following transarterial chemoembolisation in patients with unresectable hepatocellular carcinoma. Aliment Pharmacol Ther. Jun 2012;35(11):1343-50. doi:10.1111/j.1365-2036.2012.05089.x

No competing interests reported.

Supplementaryfile.docx

Download PDF

Editorial decision: Revision requested
09 Oct, 2024
Reviews received at journal
08 Oct, 2024
Reviews received at journal
04 Oct, 2024
Reviews received at journal
29 Sep, 2024
Reviews received at journal
29 Sep, 2024
Reviews received at journal
27 Sep, 2024
Reviews received at journal
24 Sep, 2024
Reviews received at journal
23 Sep, 2024
Reviewers agreed at journal
22 Sep, 2024
Reviewers agreed at journal
19 Sep, 2024
Reviewers agreed at journal
17 Sep, 2024
Reviewers agreed at journal
17 Sep, 2024
Reviewers agreed at journal
17 Sep, 2024
Reviewers agreed at journal
16 Sep, 2024
Reviewers agreed at journal
16 Sep, 2024
Reviewers agreed at journal
16 Sep, 2024
Reviewers agreed at journal
16 Sep, 2024
Reviewers agreed at journal
16 Sep, 2024
Reviews received at journal
13 Sep, 2024
Reviewers agreed at journal
13 Sep, 2024
Reviewers invited by journal
10 Sep, 2024
Editor assigned by journal
10 Sep, 2024
Editor invited by journal
29 Aug, 2024
Submission checks completed at journal
28 Aug, 2024
First submitted to journal
20 Aug, 2024

You are reading this latest preprint version

Predicting Progression-free Survival in Primary Liver Cancer Patients via Radiomics from DSA during TACE

Status:

Version 1

Abstract

Figures

Introduction

Method

Study population

Procedure of TACE

Follow-up and evaluation

Study endpoint

Image segmentation and radiomics feature extraction

Feature selection

Radiomics signatures

Risk stratification

Construction of the predictive model

Model validation

Statistical analysis

Results

Baseline information and follow-up results

Radiomics feature selection and the radiomics signature

Risk stratification on the basis of the Radscore

Factors influencing the prognosis of PLC patients after TACE

Prediction model construction and visualization

Model evaluation and validation

Discussion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1