Prediction of the efficacy of first transarterial chemoembolization for advanced hepatocellular carcinoma via a clinical-radiomics model

Abstract

BACKGROUND

Hepatocellular carcinoma (HCC) is a common tumor with a poor prognosis. Early intervention is essential; thus, good prognostic markers to identify patients who benefit from first transarterial chemoembolization (TACE) are needed.

AIM

To investigate the efficacy of computed tomography (CT) radiomics in predicting the success of the first TACE in patients with advanced HCC and to develop an early prediction model based on clinical radiomics features.

METHODS

Data from 122 patients with advanced HCC treated with TACE were analyzed. Intratumoral and peritumoral areas on arterial and venous CT images were selected to extract radiomic features, which were screened in the training cohort using the minimum redundancy maximum correlation. Then, support vector machines were used to construct the model. To construct a receiver operating characteristic curve, the predictive efficacy of each model was evaluated on the basis of the area under the curve (AUC).

RESULTS

Among the 122 patients, 72 patients were effectively treated via TACE, and in 50 patients, this treatment was ineffective. In the radiomics model, the areas under the curve of the venous phase model were 0.867 (95%CI: 0.790-0.940) in the training cohort and 0.755 (0.600-0.910) in the validation cohort, indicating good predictive efficacy. The multivariate logistic regression results indicated that preoperative alpha-fetoprotein levels (P = 0.01) were a risk factor for TACE. The screened clinical features were combined with the radiomic features to construct a combined model. This combined model had an AUC of 0.92 (0.87-0.95) in the training cohort and 0.815 (0.67-0.95) in the validation cohort.

CONCLUSION

CT radiomics has good value in predicting the efficacy of the first TACE treatment in patients with HCC. The combined model was a better tool for predicting the first TACE efficacy in patients with advanced HCC and could provide an efficient predictive tool to help with the selection of patients for TACE.

Key Words: Hepatocellular carcinoma; Transarterial chemoembolization; Radiomics; Computed tomography; Prognosis; Treatment response

Core Tip: Hepatocellular carcinoma (HCC) is often diagnosed at advanced stages, where transarterial chemoembolization (TACE) serves as a key therapy. However, nearly 50% of patients show poor TACE response due to tumor heterogeneity. This study integrates preoperative computed tomography radiomics and clinical factors to build a predictive model for TACE efficacy. Radiomics noninvasively extracts quantitative imaging features reflecting tumor pathophysiology, enabling precise assessment of treatment response. The combined model stratifies patients by predicted risk, guiding timely transition to alternative therapies (e.g., targeted drugs or immunotherapy) for non-responders. This approach enhances personalized HCC management, optimizes resource allocation, and improves survival outcomes through data-driven clinical decision-making.

INTRODUCTION

Hepatocellular carcinoma (HCC) is closely associated with chronic liver disease[1]. There is a clear geographical distribution of HCC, with nearly half of the annual new cases worldwide occurring in China[2]. HCC is highly malignant, insidious, and often asymptomatic in the early stages, with approximately 2/3 of patients being diagnosed in advanced stages of the disease without a chance of surgical treatment. HCC is characterized by multidisciplinary involvement and the coexistence of multiple treatment modalities. According to the Barcelona Clinical Liver Cancer (BCLC) staging system and the European and United States guidelines for the treatment of HCC, local treatment via TACE has become an important option for advanced liver cancer[3-5]. However, the prognosis of HCC patients after TACE can differ considerably because of tumor heterogeneity, which involves factors such as tumor burden, vascular invasion, and liver function[6]. Nearly half of patients with HCC do not respond to TACE treatment and have a poor prognosis[7]. For patients who do not respond to TACE, timely conversion to targeted drugs (sorafenib, lenvatinib, or donafenib) or immunotherapy can prevent further liver dysfunction and prolong overall survival (OS). How to select an accurate treatment plan for HCC patients is an important part of precision HCC therapy, and it is also an urgent clinical problem. Computed tomography (CT) is an important imaging method for HCC screening because of its short examination time and low cost[8]. Radiomics refers to the conversion of digital medical images into a myriad of quantitative features to provide information about the pathophysiology of tumors. Radiomics plays important roles in tumor detection, diagnosis, prognostic assessment, prediction of treatment response, and follow-up by analyzing a large number of images (CT and magnetic resonance imaging) to extract valuable radiomic features; achieve tumor segmentation, feature extraction and model building; and evaluate various features at the morphological, cellular-molecular, and genetic levels of tumors[9,10]. Radiomics is widely used as a noninvasive, reproducible and simple research method. The appropriate integration of radiomic features and clinical factors can improve accuracy in complex clinical decision-making. In this study, preoperative CT-enhanced images of patients and clinical risk factors were used to construct a combined model for predicting TACE efficacy, to judge response on the basis of patients' predicted risk values, and to provide a reference for individualized treatment.

MATERIALS AND METHODS

Study population

Data from HCC patients who underwent TACE treatment in the interventional department of our hospital between January 2014 and December 2020 and fulfilled the following criteria were retrospectively analyzed. The inclusion criteria were as follows: (1) Patients diagnosed with HCC; (2) BCLC stage B to C; (3) At least one TACE; (4) Complete follow-up data; (5) CT-enhanced scan performed within 1 week prior to the initial TACE treatment; and (6) Imaging follow-up within 3 months after the procedure. The exclusion criteria were as follows: (1) CT image artifacts affecting lesion observation and analysis; (2) Combination with other cancers; (3) Incomplete data; (4) Other treatments (radiotherapy, chemotherapy, targeted drug therapy, and immunotherapy) prior to TACE; and (5) Liver or renal failure. The final number of eligible patients was 122, comprising 106 men and 16 women aged 23–80 years, with a mean age of 55 ± 10 years. Patients were randomly divided into a training cohort and a validation cohort on the basis of a random split-sample (7:3) approach. The model was built with data from the training cohort and validated with data from the validation cohort.

Devices

A Siemens Sensation 16-layer spiral CT machine or Siemens Somatom Definition Flash dual-source CT machine was used for scanning and enhancement scans with the following parameters: Layer thickness and layer interval of 5-mm, tube voltage of 120 kV, and tube current of 250 mA. For enhanced scans, 70-80 mL of iohexol (300 mg/100 mL) was injected at a flow rate of 3.0 mL/s via a high-pressure syringe in the elbow vein, and dynamic scans of the arterial and venous phases were performed 25-35 s and 60-75 s after injection via standard algorithms for reconstruction.

Treatment

All patients were treated with conventional TACE, and the procedure was performed by interventionalists with more than 10 years of experience. The patient was placed supine on a digital subtraction angiography bed, and after local anesthesia was in effect, a modified Seldinger's puncture technique was used to access the tumor via the femoral artery, and the arterial sheath, catheter, and guidewire were inserted in turn. A 2.5-F microcatheter was then used to super-selectively cannulate the blood supply vessels of the liver segment or subsegment for slow infusion of chemotherapeutic agents. The TACE procedure was a classic "sandwich" approach, with a slow infusion of pirarubicin, oxaliplatin, and 5-fluorouracil through the catheter followed by a slow infusion of an emulsion of pirarubicin and iodinated oil, followed by embolization of the tumor vessels with gelatin sponge pellets until the contrast agent was stopped. The embolization endpoint was the disappearance of tumor staining.

Treatment response

Treatment efficacy was assessed according to the modified Response Evaluation Criteria in Solid Tumors criteria[11]: Complete response (CR): CT-enhanced scans show no enhancement in the arterial phase in all target lesions. Partial response (PR): A 30% reduction in the sum of the diameters of the tumor lesions (arterial phase). Progressive disease (PD): The total diameter of the target lesions (arterial phase) increased by 20%, or new lesions appeared. Stable disease (SD): The total diameter of the target lesion (arterial phase) did not decrease as much as it did in the PR and did not increase as much as it did in the PD. CR and PR were defined as effective treatments and PD and SD were defined as ineffective.

Laboratory data and imaging features

All patients' laboratory tests were completed within 1 week prior to the procedure. Clinical information included age, sex, and history of hepatitis (hepatitis B, hepatitis C, alcoholic hepatitis). The laboratory data included alpha-fetoprotein (AFP), alanine aminotransferase, aspartate aminotransferase, albumin, total bilirubin, prothrombin time, and Child-Pugh classification. Imaging features included lesion location, number, diameter (diameter of the largest lesion in multinodular HCCs), presence or absence of an envelope, presence or absence of necrosis, presence or absence of intratumoral hemorrhage, tumor rupture, and portal vein thrombosis. The tumor envelope was defined as a smooth, uniform, dense shadow around the tumor in the venous phase. Intratumoral necrosis was defined as a hypointense shadow without enhancement in all stages of the tumor. The location of the tumor was categorized as the left, right, or caudate lobe of the liver. The number of tumors was classified as single or multiple. To determine tumor size, the maximum axial length diameter of the tumor, including the envelope, was measured on arterial phase images. Intratumoral hemorrhage was defined as a high density on plain scan without reaching the density of a calcification. Portal vein carcinoma thrombosis is characterized by a dilated portal or hepatic vein with a filling defect and perihepatic disorganized collateral circulation upon enhancement[12].

Segmentation of regions of interest and extraction of radiomic features

Each layer of the images was segmented by the open-source software Slicer (Version 4.10.2 https://www.slicer.org/), which was used to segment the tumor and the peritumoral 5-mm and 10-mm regions of interest (ROIs) (Figure 1). If the tumor was large and the perimeter of the tumor extended beyond the edge of the liver, it was sufficient to segment the maximum distance from the tumor to the edge of the liver. All image segmentations were performed by two radiologists experienced in diagnosing HCC (with 5 and 10 years of work experience). The radiologists needed to be made aware of the patient's clinical information and survival status. The ROI was first manually segmented by a radiologist with 5 years of experience and then confirmed by a radiologist with 10 years of experience; If there are discrepancies, consult a third senior radiologist and make necessary adjustments. Radiomics features were extracted via the Slicer Radiomics package in Slicer, and a total of 1218 radiomic features were extracted, including first-order statistics, shape features, gray level co-occurrence matrix features, gray level run length matrix features, and gray level size zone matrix features. The features extracted in this study are in accordance with the recommendations of the Imaging Biomarker Standardization Initiative (IBSI) for the definition of features[13].

Figure 1 Schematic of the segmentation of the tumor region of interest. A: Original tumor image; B: Tumor regions of interest (ROI); C: Peritumoral 10-mm ROI; D: 3D simulation of the tumor; E: 3D simulation of the tumor and tumor peritumoral regions.

Feature screening and model building

For the clinical model, the clinical and imaging features that differed (P ≤ 0.1) were subjected to multifactorial logistic regression analysis. The clinical and imaging features selected at P < 0.05 were used to construct a clinical model via a support vector machine (SVM) approach. For the radiomic model, radiomic features were downscaled and filtered by Z score normalization. The outlier median substitution, intragroup consistency test (ICC), minimum redundancy maximum relevance (mRMR), and least absolute shrinkage and selection operator (LASSO) were used to select the most promising radiomic features and calculate the radiomic score (Rad-score) for each patient. The SVM algorithm was used to construct models for each of the arterial, venous, and arterial + venous phases, and the best model was selected to construct the 5-mm and 10-mm peritumoral models. The final screening identified the radiomics model that predicted the optimal efficacy of TACE for HCC. For the combined model, the clinical and imaging features screened for the clinical model were used in conjunction with the Rad score to construct the combined model.

Receiver operating characteristic curve

A receiver operating characteristic (ROC) curve was drawn. The predictive efficacy of each model was evaluated based on the area under the curve (AUC), and the data were verified in the verification cohort. To make the model more intuitive, the combined model is presented as a visual nomogram. The model's predictive performance was assessed via calibration curves, and the fit of the model was evaluated via the Hosmer-Lemeshow test. Moreover, we employed the Delong test to compare the different ROC curves. Decision curve analysis was used to determine the usefulness of the model.

Statistical analysis

SPSS (Version 18.0) and R software (Version 3.6.3) were used for statistical analysis and model building. The normality of the quantitative data was assessed via the Kolmogorov–Smirnov test. Continuous variables satisfying the normal distribution criteria were evaluated via the independent samples t-test; otherwise, the Mann-Whitney test was used. Categorical variables were tested via the χ² test or Fisher's test.

RESULTS

Demographic and clinical characteristics of the patients

A total of 122 patients were enrolled in this study. Among these patients, 72 patients were treated effectively, with a rate of 59%, whereas 50 patients were treated ineffectively, with a rate of 41%. The study population was randomly divided into a training cohort (n = 85) and a validation cohort (n = 37). In the training and validation cohorts, the clinical data and imaging features were not significantly different, except for a history of viral hepatitis B (P = 0.04; Tables 1 and 2). Data from the training cohort were statistically analyzed (Tables 3 and 4), and characteristics with P ≤ 0.1 were included in the multivariate regression analysis. Finally, the prothrombin time (P = 0.090), AFP level (P = 0.003), presence of an envelope (P = 0.100), and presence of portal vein carcinoma thrombus (P = 0.073) were included in the multifactorial regression analysis. The results revealed that the difference in the AFP level was statistically significant (P < 0.05; Table 5), and the AFP level was eventually included in both the clinical model and the combined model.

Table 1 Clinical information of the training and validation cohorts, n (%)/mean ± SD.

Characteristics	Training cohort (n = 85)	Validation cohort (n = 37)	P value
Sex			0.430
Male	73 (85.9)	33 (89.2)
Female	12 (14.1)	4 (10.8)
Age	54.35 ± 10.930	57.43 ± 10.219	0.147
Child-Pugh class			0.213
A	67 (78.8)	26 (70.3)
B	18 (21.2)	11 (29.7)
Hepatitis			0.040
None	4 (4.7)	1 (2.7)
Hepatitis B	79 (92.9)	32 (86.5)
Hepatitis C	1 (1.2)	1 (2.7)
Alcoholic hepatitis	1 (1.2)	3 (8.5)
PT (s)			0.458
≤ 14	80 (94.1)	34 (91.9)
> 14	5 (5.9)	3 (8.1)
TB (μmol/L)			0.173
≤ 17.1	41 (48.2)	22 (59.5)
> 17.1	44 (51.8)	15 (40.5)
ALB (g/L)			0.144
≤ 35	45 (52.9)	15 (40.5)
> 35	40 (47.1)	22 (59.5)
AST (U/L)			0.357
≤ 40	18 (21.2)	6 (16.2)
> 40	67 (78.8)	31 (83.8)
ALT (U/L)			0.194
≤ 50	59 (69.4)	22 (59.5)
> 50	26 (30.6)	15 (40.5)
AFP (ng/mL)			0.326
≤ 400	27 (31.8)	14 (37.8)
> 400	58 (68.2)	23 (62.2)

PT: Prothrombin time; TB: Total bilirubin; ALB: Albumin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; AFP: Alpha-fetoprotein.

Table 2 Imaging features of the training and validation cohorts, n (%)/mean ± SD.

Characteristics	Training cohort (n = 85)	Validation cohort (n = 37)	P value
Diameter (mm)	95.04 ± 32.988	92.97 ± 35.158	0.752
Number of tumors			0.351
One	37 (43.5)	14 (37.8)
Multiple	48 (56.5)	23 (62.2)
Location			0.516
Right	59 (69.4)	25 (67.6)
Left	23 (27.1)	12 (62.4)
Caudate lobe	3 (3.5)	0 (0)
Envelope			0.409
Absent	31 (36.5)	15 (40.5)
Present	54 (63.5)	22 (59.5)
Tumor necrosis			0.279
Absent	17 (20.0)	5 (13.5)
Present	68 (80.0)	32 (86.5)
Intratumoral hemorrhage			0.598
Absent	76 (89.4)	33 (89.2)
Present	9 (10.6)	4 (10.8)
Tumor rupture			0.074
Absent	78 (91.8)	37 (100.0)
Present	7 (8.2)	0 (0)
Portal vein thrombosis			0.129
Absent	39 (45.9)	11 (29.7)
Present	46 (54.1)	26 (70.3)

Table 3 Relationships between the clinical characteristics of the training cohort and the efficacy of transarterial chemoembolization treatment, n (%)/mean ± SD.

Characteristics	Effective (n = 50)	Invalid (n = 35)	P value
Sex			0.36
Male	44 (82.9)	29 (82.9)
Female	6 (17.1)	6 (17.1)
Age	55.74 ± 10.764	52.37 ± 11.014	0.163
Child-Pugh class			0.523
A	39 (78.0)	28 (80.0)
B	11 (22.0)	7 (20.0)
Hepatitis			0.511
None	2 (4)	2 (18.4)
Hepatitis B	47 (94)	32 (26.3)
Hepatitis C	1 (2)	0 (31.6)
Alcoholic hepatitis	0 (0)	1 (23.7)
PT(s)			0.090
≤ 14	49 (98.0)	31 (88.6)
> 14	1 (2.0)	4 (11.4)
TB (μmol/L)			0.567
≤ 17.1	24 (48.0)	17 (48.6)
> 17.1	26 (52.0)	18 (51.4)
ALB (g/L)			0.192
≤ 35	24 (48.0)	21 (60.0)
> 35	26 (52.0)	14 (40.0)
AST (U/L)			0.151
≤ 40	13 (26.0)	5 (14.3)
> 40	37 (74.0)	30 (85.7)
ALT (U/L)			0.463
≤ 50	34 (68.0)	25 (71.4)
> 50	16 (32.0)	10 (28.6)
AFP (ng/mL)			0.003
≤ 400	22 (44.0)	5 (14.3)
> 400	28 (56.0)	30 (85.7)

PT: Prothrombin time; TB: Total bilirubin; ALB: Albumin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; AFP: Alpha-fetoprotein.

Table 4 Relationships between the imaging characteristics of the training cohort and the efficacy of transarterial chemoembolization treatment, n (%)/mean ± SD.

Characteristics	Effective (n = 50)	Invalid (n = 35)	P value
Diameter (mm)	94.42 ± 33.742	95.91 ± 32.347	0.839
Number of tumors			0.157
One	19 (38.0)	18 (51.4)
Multiple	31 (62.0)	17 (48.6)
Location			0.567
Right	35 (70.0)	24 (68.6)
Left	13 (26.0)	10 (28.6)
Caudate lobe	2 (4.0)	1 (2.9)
Envelope			0.100
Absent	15 (30.0)	16 (45.7)
Present	35 (70.0)	19 (54.3)
Tumor necrosis			0.395
Absent	11 (22.0)	6 (17.1)
Present	39 (78.0)	29 (82.9)
Intratumoral hemorrhage			0.551
Absent	45 (90.0)	31 (88.6)
Present	5 (10.0)	4 (11.4)
Tumor rupture			0.612
Absent	46 (92.0)	32 (91.4)
Present	4 (8.0)	3 (8.6)
Portal vein thrombosis			0.073
Absent	26 (52.0)	13 (37.1)
Present	24 (48.0)	22 (62.9)

Table 5 Results of the multifactorial logistic regression analysis of the efficacy of transarterial chemoembolization treatment.

Characteristics	OR	95%CI	P value
Portal vein thrombosis	4.906	0.662-36.352	0.421
AFP	4.710	1.438-15.420	0.010
PT	7.308	0.707-75.488	0.095
Envelope	0.701	0.223-2.206	0.543

OR: Odds ratio; PT: Prothrombin time; AFP: Alpha-fetoprotein.

Radiomics model

On the basis of the CT-enhanced arterial and venous images, a total of 1218 radiomics features were extracted for each lesion, with 840 features exhibiting an ICC ≥ 0.8 retained for further analysis. and the features were subjected to Z score normalization and outlier median substitution, mRMR, and LASSO (Figure 2) to reduce dimensionality and filtering. The SVM algorithm was subsequently used to construct radiomic models for the arterial, venous, and arterial + venous phases of the tumors in the training cohort. The corresponding AUC values were 0.885 (95%CI: 0.814-0.955), 0.867 (95%CI: 0.790-0.940), and 0.840 (95%CI: 0.758-0.923), respectively, and were independently validated on the validation cohort, with corresponding AUC values of 0.592 (95%CI: 0.400-0.783), 0.755 (95%CI: 0.600-0.910), and 0.645 (95%CI: 0.461-0.829), respectively. Among these models, the tumor venous phase model had excellent efficacy in both the training and validation cohorts, on the basis of which the peritumoral 5-mm and 10-mm models were built. The AUC values for the peritumoral 5-mm model in the training and validation cohorts were 0.866 (95%CI: 0.791-0.941) and 0.594 (95%CI: 0.402-0.786), respectively, whereas the AUC values for the peritumoral 10-mm model were 0.922 (95%CI: 0.866-0.977) and 0.603 (95%CI: 0.408-0.799), respectively. Although both models performed well in the training cohort, they performed poorly in the validation cohort (Table 6). Ten radiomic features were ultimately extracted, including four first-order features (wavelet-HHL-first-order-range, wavelet-LLL-first-order- skewness, wavelet-HLH-first-order-skewness, and wavelet-LLL-GLcm-maximum probability), five gray-level matrix features (wavelet-LHL-glcm-maximum probability, wavelet-HHL-glcm-Imc1, wavelet-LLH-glcm-MCC, wavelet-LHL-glcm -Imc1, and wavelet wavelet-LLL-Glcm-maximum probability), and one gray-level run-length matrix feature (wavelet-HHH-Glrlm-LongRun Empasis). The final radiomic model was the venous phase model.

Figure 2  Schematic diagram of Lasso screening.

Table 6 Comparison between the training and validation cohorts of the radiomics models.

Model	Training cohort (n = 85)				Validation cohort (n = 37)
	AUC	95%CI	Sensitivity%	Specificity%	AUC	95%CI	Sensitivity%	Specificity%
Arterial phase	0.885	0.814-0.955	0.829	0.613	0.592	0.400-0.783	0.476	0.428
Venous phase	0.867	0.790-0.940	0.863	0.686	0.755	0.600-0.910	0.818	0.533
Arteria + venous phase	0.84	0.758-0.923	0.823	0.657	0.645	0.461-0.829	0.590	0.467
Peritumor 5- mm (venous phase)	0.866	0.791-0.941	0.863	0.656	0.594	0.402-0.786	0.591	0.667
Peritumor 10- mm (venous phase)	0.922	0.866-0.977	0.921	0.686	0.603	0.408-0.799	0.682	0.467

AUC: Area under the curve.

Combined model

The screened clinical feature (AFP level) and Rad score were used to construct a combined model using SVMs. The ROC curve (Figure 3) was used to analyze the predictive performance of each model and to calculate the AUC, accuracy, sensitivity, and specificity. Each model was then independently validated in the validation cohort, and a nomogram was built (Figure 4). The AUCs for the clinical model, radiomics model, and combined model in the training cohort were 0.65 (95%CI: 0.56-0.74), 0.86 (95%CI: 0.79-0.94), and 0.92 (95%CI: 0.87-0.95), respectively, whereas in the validation cohort, they were 0.60 (95%CI: 0.44-0.76), 0.75 (95%CI: 0.60-0.91), and 0.815 (95%CI: 0.67-0.95), respectively. In both the training and validation cohorts, the AUC values of the combined model were greater than those of the clinical model and the radiomic model; i.e., the combined model was optimal for predicting the efficacy of TACE treatment. However, when the radiomics model was compared with the clinical-radiomics model using the DeLong test in the training and validation sets, respectively, no statistically significant difference in predictive performance was observed (Table 7).The predictive accuracy of the nomogram was corrected via the calibration curve (Figure 5), from which it is evident that the goodness of fit is good, and it can be concluded that there is good agreement between the nomogram-predicted probability of efficacy of TACE treatment and the actual probability of efficacy of TACE treatment. The results of the decision curve analysis (Figure 6) indicate that the combined model predicts greater net benefit than the clinical model for postoperative TACE outcomes over the vast majority of the threshold probability range.

Figure 3 Receiver operating characteristic curves. A: Receiver operating characteristic (ROC) curves of different models in the training cohorts; B: ROC curves of different models in the validation cohorts.

Figure 4 Nomogram. A nomogram prediction model of first-line treatment efficacy in patients with hepatocellular carcinoma consisting of two predictors, the alpha-fetoprotein level and the Rad-score, was developed.

Figure 5  Nomogram calibration curve.

Figure 6 Decision curve analysis. The results of the decision curve analysis revealed that the net clinical benefit of the combined model was greater than that of the clinical model at all threshold probabilities between 0.16 and 0.76.

Table 7 Comparison of the performance of different predictive models using the DeLong test in the training and validation cohorts.

Cohort	Model comparison	P value
Training cohort	Clinical model vs radiomics model	< 0.023
	Clinical model vs combined model	< 0.015
	Radiomics model vs combined model	0.453
Validation cohort	Clinical model vs radiomics model	< 0.036
	Clinical model vs combined model	< 0.021
	Radiomics model vs combined model	0.632

DISCUSSION

HCC is associated with high morbidity, poor treatment efficacy, and high mortality. Local treatment, represented by TACE, has become an important approach for intermediate and advanced HCC. Substantial differences in the prognosis of HCC patients treated with TACE exist due to tumor heterogeneity in terms of tumor load, vascular invasion, liver function, and other factors[6,14]. Nearly half of the patients with HCC do not respond to TACE treatment. In this study, 122 patients chose conventional TACE, and 72 were treated effectively, with a rate of 59%, similar to that reported in the literature[7]. Patients who do not respond to TACE therapy are promptly treated with physical therapy (microwave ablation, radiofrequency ablation, cryoablation), targeted drugs (sorafenib, lenvatinib, or donafenib), or immunotherapy (atelelizumab in combination with bevacizumab or sindilizumab in combination with bevacizumab) to control tumor growth, prevent further liver damage, and prolong OS[15-17].

Radiomics, first proposed by Dutch researchers in 2012, refers to the conversion of digital medical images into a myriad of quantitative features to provide information about the pathophysiology of tumors. It plays an important role in tumor detection, diagnosis, prognostic assessment, prediction of treatment response, and follow-up[9,18,19]. However, the clinical application of radiomics has been hampered by the need for standardized guidelines. Recently, some investigators have advocated the use of radiology software tools that conform to the IBSI, aiming to standardize the extraction of histological imaging features[20].

The present study retrospectively analyzed the clinical data of 122 patients via multifactorial logistic regression analysis and revealed that AFP was an independent clinical risk factor for the efficacy of TACE treatment, as patients with AFP ≤ 400 ng/mL had better treatment efficacy than those with AFP > 400 ng/mL. Chen et al[21]. reported that AFP levels could be used as a predictor for assessing tumor response prior to TACE and concluded that patients with high AFP levels (> 200 ng/mL) had a lower response rate to TACE than those with low AFP levels. AFP is a glycoprotein primarily produced during fetal development by the liver and yolk sac. In adults, elevated AFP levels are strongly associated with HCC, AFP is not merely a bystander biomarker but an active player in HCC progression through immune evasion, anti-apoptosis, and pro-angiogenic mechanisms. Increased AFP values are associated with lower patient survival, higher rates of tumor recurrence, and poor prognosis for patients with advanced HCC[22-24]. This result is consistent with what has been reported previously. Tumor location (segments I and IV) has been reported in the national and international literature to be associated with poor efficacy of TACE treatment. Segments I and IV originate from branch arteries of the left and right hepatic arteries with multiple vascular anastomotic networks, providing partial or total blood supply to the tumor and resulting in reduced TACE efficacy or even tumor recurrence and metastasis[25,26]. The study population comprised patients with BCLC stage C disease. Most of these patients had large masses not limited to a particular segment, often involving multiple lobes or segments, and the larger the tumor was, the more arteries involved in the blood supply and the more complex the vascular anastomosis, resulting in poorer outcomes.

The radiomic features extracted in this study were derived mainly from the venous phase and may be related to tumor heterogeneity. Tumor heterogeneity is closely related to tumor prognosis, and increased heterogeneity in the arterial phase can predict tumor recurrence after HCC treatment. In advanced HCC, enhancement is less pronounced in the arterial phase than in the venous phase due to poor tumor perfusion, poor vascular function, and reduced arterial blood flow, which is associated with high expression of HCC-specific markers[27]. For example, it is associated with upregulated vascular endothelial growth factor A (VEGFA) and fibroblast growth factor receptor 4 levels[28,29]. This relationship can be explained by tumor progression, which leads to a reduction in arterial blood flow, as higher intercellular pressure causes the arterial capillaries to close.

Moreover, hypoxia-induced VEGFA expression is negatively correlated with perfusion, which may be more pronounced in tumors with inadequate blood supply. Hypovascular HCC has a poor arterial supply, is difficult to control with conventional TACE, and has poor therapeutic efficacy[30]. A number of antiangiogenic drugs have been developed as a result, reflecting the prognostic significance of the radiomic features, as well as the fact that advanced HCC shows greater tumor heterogeneity in the venous phase. The AUC of the venous phase model in our study revealed excellent predictive efficacy in both the training and validation cohorts.

Single morphological features are of limited value in predicting the response to tumor therapy, and studies have shown that GLCM features represent heterogeneity in peripheral tumor regions. In other words, the wavelet transform can provide a full range of spatial and frequency distributions for characterizing tumor interiors and peripheral tumor regions on the basis of low- and high-frequency signals; these features can improve the performance of radiomic models[29,31]. In this study, the CT-based histological features consisted mainly of wavelet transform textures and were mostly GLCM features. The wavelet transform decomposes the image into high- and low-frequency signals in the intratumoral and peritumoral regions, which may be a powerful indicator of the heterogeneity of the tumor microenvironment, but further genomic and histopathological data are needed to verify. Zhou et al[32] demonstrated that wavelet-based features can be used for disease diagnosis and treatment response prediction. These features may further reflect the properties of the tumor and its periphery in terms of spatial heterogeneity in multiple dimensions. Radiomics features can provide more detailed information about tumor biology, as well as the tumor microenvironment, and are complementary to images alone.

Our study established 5-mm and 10-mm peritumoral models on the basis of the venous phase. Although both models performed well in the training cohort, they performed poorly in the validation cohort, related to the lack of specificity of peritumoral monocytes in promoting the upregulation of autophagy in tumor cells by the secretion of tumor necrosis factor α and interleukin-1β in the invasive marginal regions of HCC, leading to an increase in inflammatory cells[33]. The number of peritumoral hepatic macrophages increased with decreased expression of nitric oxide synthase 2 and increased human leukocyte antigen DR isotype-positive cells, the latter of which upregulate T lymphocytes by producing multiple cytokines and chemokines. The humoral cytokine response observed in the peritumoral region may be associated with HCC venous metastasis[34]. However, owing to the heterogeneity of peritumoral tissues and the lack of algorithmic standardization, radiomic features extracted from peritumoral tissue are less important than those extracted from the tumor fraction[35]. Some studies have shown that the size of the peritumoral region varies on the basis of biological and clinical factors and that it is not appropriate to obtain the peritumoral region by setting a consistent distance from the tumor[21]. In this study, 5-mm and 10-mm peritumoral histological imaging models were developed; however, the predictive efficacy of both models was unsatisfactory, related to the following factors: (1) Individual peritumoral features of advanced HCC are not specific; (2) Advanced HCC tumors invade adjacent liver tissues, resulting in unclear boundaries and a lack of standardization of peritumoral target area segmentation; and (3) The sample size was insufficient and should be expanded in further studies.

Nomograms are used as statistical models to explain the proportion of multiple risk factors by assigning a total number of points to each patient. It requires input of routine clinical and imaging data (e.g., AFP levels, Child-Pugh scores, and standard radiomic features), ensuring minimal disruption to current workflows. A user-friendly interface prototype is under development to simplify input/output interactions for clinicians (e.g., risk probability visualization and actionable alerts). We emphasize that the model is not intended to replace clinical judgment but to augment risk stratification. For example, high-risk predictions could prioritize patients for closer surveillance or adjuvant therapies. Various studies have shown that radiomics nomograms can effectively and comprehensively predict postoperative patient outcomes[36]. Compared with the radiomics nomogram alone, the clinical-radiomics nomogram achieved better prognostic performance, with a higher C-index and better calibration. Decision curve analysis revealed that the clinical-radiomic nomogram outperformed the radiomic nomogram for most reasonable threshold probability ranges.

In the present study, a combined clinical-radiomic model was constructed and validated for predicting the efficacy of the response to the first TACE treatment in patients with intermediate-to-advanced HCC. The model consists mainly of 10 venous-phase radiomic features and one clinical feature. The combined model exhibits superior predictive efficacy and clinical application, providing clinicians with a visual, individualized efficacy prediction tool in the form of a nomogram. It is helpful to predict the early curative effect of TACE before treatment and provides an important reference for TACE treatments of patients with HCC. For patients with predicted ineffective TACE treatment, other options can be promptly selected to prevent further liver damage, reduce the economic burden and unnecessary suffering of patients, and improve their survival time and quality of life.

This study was a retrospective study, which was prone to selection bias and lacked externally validated data. The sample size was small, manual ROI segmentation was time-consuming, and individualized differences in the segmentation ROI existed. Most of the patients in this study had hepatitis B, which may have a certain population selection bias. But we are actively partnering with three institutions to validate this model in a prospective cohort of > 300 patients to evaluate clinical utility, cost-effectiveness, and provider acceptance. Moreover, we did not consider the radiation dose or drug dosage because the vascular conditions during the procedure differ among patients, as does the procedure duration, leading to differences in drug dosages. Despite this, our study exhibited the potential of the radiomics model, as a non-invasive and accurate method, to predict the response to TACE therapy in HCC. It is anticipated that more automated deep learning tools will be developed for future work to minimize human errors. In future studies, we will record the radiation dose and specific drug dosage for each patient, increase the sample size, collect multicenter data, and perform external validation.

CONCLUSION

Radiomics has better performance in predicting the treatment efficacy of the first TACE in patients with HCC. Compared with the clinical model alone and the radiomic model, the combined model has better predictive efficacy, which is beneficial for clinical decision-making.