Interpretable radiomics model based on magnetic resonance imaging to predict responses to transarterial chemoembolization for hepatocellular carcinoma

Abstract

BACKGROUND

Transarterial chemoembolization (TACE) is important role for the treatment of unresectable hepatocellular carcinoma (HCC). However, owing to the heterogeneity of HCC tumors, TACE efficacy varies among individual HCC patients. Accurate preoperative prediction of responses to TACE among HCC patients could guide the development of individualized treatment strategies and improve patient outcomes.

AIM

To investigate the predictive value of multiple-sequence magnetic resonance imaging (MRI) radiomic features combined with clinical indices for the response to TACE among HCC patients and to develop an interpretable machine learning model.

METHODS

A total of 116 patients with HCC who underwent TACE were retrospectively enrolled. Patients in this study were randomly divided into a training group and a validation group at a ratio of 7:3. The response to TACE was evaluated according to the modified Response Evaluation Criteria in Solid Tumors. Radiomic features were extracted from axial fat-suppressed T2-weighted imaging, arterial phase and portal venous phase axial dynamic contrast-enhanced MRI sequences. The least absolute shrinkage and selection operator method was used to select the best radiomic features. Univariate and multivariate logistic regression were used to select clinical predictive factors that affect the response to TACE among patients with HCC. Logistic regression was used to construct a radiomic model of each sequence, a multiple-sequence MRI radiomics model, and a radiomic-clinical (RC) model. The predictive performance of each model was evaluated using the area under the receiver operating characteristic curve. The risk factors in the RC model were subsequently interpreted via SHapley Additive exPlanations analysis.

RESULTS

The area under the receiver operating characteristic curve values of the fat-suppressed T2-weighted imaging, arterial-phase, portal-venous-phase, joint-radiomic model and RC model were 0.771, 0.668, 0.725, 0.845 and 0.929, respectively, in the training group and 0.705, 0.666, 0.675, 0.799 and 0.815, respectively, in the validation group. The clinical-radiomic model had the best predictive performance. The SHapley Additive exPlanations algorithm was used to illustrate the contribution of each feature in the RC model.

CONCLUSION

The interpretable RC model could successfully stratify HCC patients into TACE responders and TACE nonresponders and serve as a potential tool to identify more appropriate HCC patients for TACE, thus sparing patients from receiving ineffective and unnecessary treatments.

Key Words: Magnetic resonance imaging; Radiomics; Hepatocellular carcinoma; Transarterial chemoembolization; Treatment response

Core Tip: Transarterial chemoembolization (TACE) plays an important role in the treatment of unresectable hepatocellular carcinoma (HCC). However, owing to the heterogeneity of HCC tumors, TACE efficacy varies among individual HCC patients. Accurate preoperative prediction of responses to TACE among HCC patients could guide the development of individualized treatment strategies and improve patient outcomes. This study aimed to investigate the predictive value of multiple-sequence magnetic resonance imaging radiomic features combined with clinical indices for the response to TACE among HCC patients and to develop an interpretable machine learning model.

INTRODUCTION

Hepatocellular carcinoma (HCC) is the sixth most common cancer and the third most common cause of cancer-related death worldwide[1,2]. Transarterial chemoembolization (TACE) is the standard treatment for Barcelona Clinic Liver Cancer stage B HCC[3,4]. Owing to the heterogeneity of HCC tumors, TACE efficacy varies among individual HCC patients[5,6]. Therefore, accurate preoperative prediction of responses to TACE among HCC patients could guide the development of individualized treatment strategies and improve patient outcomes.

Radiomics is an image analysis technology that has emerged in recent years[7-11]. Radiomics can classify and provide a prognostic analysis of diseases by mining high-throughput quantitative radiomic features from existing medical images and building models[12-16]. Magnetic resonance imaging (MRI) is characterized by the absence of radiation, high-resolution soft tissue, and multiparameter and multidirectional imaging; therefore, MRI has been recommended as a routine evaluation method for HCC TACE treatment[17]. Previous studies have shown that MRI-based radiomic models are valuable for both differentiating pathological classification[18] and predicting the microvascular invasion status[19], gene expression[20], treatment response[21-23], and prognosis of HCC patients[24-27]. Recently, in models developed by combining radiomics with machine learning, SHapley Additive exPlanations (SHAP) analysis has been applied to increase the interpretability of the models. The SHAP algorithm, which is rooted in game theory, provides a quantitative assessment of the influence of individual features on the model's predictive output. This approach facilitates a more comprehensive understanding of the model's decision-making process for clinicians, thereby fostering greater trust in the predictive accuracy and reliability of the model.

However, few studies have evaluated the value of interpretable machine learning models using multiple-sequence MRI radiomic features combined with relevant clinical risk factors to predict the treatment response to TACE among HCC patients. In this study, the ability of preoperative magnetic resonance fat-saturation (FS) T2-weighted imaging (T2WI) and dynamic contrast-enhanced (DCE) arterial phase (AP) and portal venous phase (PVP) radiomic features combined with related clinical risk factors to predict the response to TACE among patients with HCC were explored. Furthermore, the SHAP algorithm was employed to quantify the contribution of each risk factor to the model. We assume that the interpretable radiomic-clinical (RC) model could serve as a potential tool to identify more appropriate HCC patients for TACE, thus sparing patients from receiving ineffective and unnecessary treatments.

MATERIALS AND METHODS

Clinical data

This study was approved by the Ethics Committee of the Affiliated Hospital of North Sichuan Medical College (Approval No. 2025ER8-1) and was performed in accordance with the Declaration of Helsinki. Because this study was a retrospective study with anonymous data collection, the requirement for informed consent was waived. This study included 116 HCC patients who underwent TACE at our hospital, including 104 males and 12 females. The inclusion criteria were as follows: (1) Patients were pathologically confirmed to have HCC or met the diagnostic criteria for HCC according to the 2018 Practice Guidance by the American Association for the Study of Liver Diseases[19]; (2) Contrast-enhanced MRI (CE-MRI) examination was performed within 1 month before TACE, and follow-up was performed within 3 months after TACE to evaluate treatment efficacy; and (3) The Child-Pugh classification of liver function was A or B. The exclusion criteria were as follows: (1) Radiotherapy and chemotherapy intervention were performed before TACE treatment; (2) Patients with incomplete clinical data; (3) Poor image quality that affected observation of the lesion; and (4) Patients with other types of tumors.

MRI scan

All patients were subjected to MRI examination before TACE with a Discovery 750 3.0T magnetic resonance system from GE, United States. The patients fasted for 4 hours before examination. Patients were trained for the breathing exercise, i.e., holding their breath at the end of expiration. The MRI scans were performed in the supine position. The scan sequences included FS-T2WI and axial CE-MRI sequences in the AP and PVP. A high-pressure syringe was used to inject the contrast agent gadopentetate dimeglumine into the dorsal vein of the hand at a dose of 15-20 mL and a rate of 2-2.5 mL/second.

TACE operation

Routine disinfection and draping were performed. After local anesthesia, the right femoral artery was punctured, and a 5F vascular sheath was inserted into the femoral artery through a guidewire. Afterward, under the guidance of a 0.035-inch guidewire, the RH catheter was inserted into the hepatic artery for digital subtraction angiography to evaluate the blood supply to the lesion. The patient was subsequently treated with chemotherapy by perfusion consisting of 100 mL of normal saline and 0.75 g of fluorouracil. The catheter was superselectively inserted into the tumor artery blood supply, and a mixed solution of iodized oil (5-20 mL), lobaplatin (30-40 mg), and epirubicin (20-30 mg) was slowly injected for chemoembolization, followed by the injection of gelatin sponge particles (the dose was determined on the basis of liver function and lesion volume).

Response evaluation

The MRI and clinical data of the patients were collected. The response to TACE was evaluated according to the modified Response Evaluation Criteria in Solid Tumors[28]. Patients exhibiting complete remission or partial remission were classified into the TACE response group, and patients exhibiting stable disease or progressive disease were classified into the nonresponse group. The patients were randomly assigned to the training group or validation group at a ratio of 7:3.

Radiomics workflow

In this study, three-dimensional Slicer 5.5.0 software (https://download.slicer.org) was used, manual segmentation was applied to sketch the volume of interest layer by layer on the magnetic resonance FS-T2WI and DCE AP and PVP images, and the corresponding magnetic resonance radiomic features were extracted (Figure 1).

Figure 1 Region of interests along the edge of the lesion. A: Fat-saturation T2-weighted imaging; B: Arterial phase image; C: Portal venous phase image; D: Three-dimensional image of the lesion.

The intraobserver and interobserver consistency of the features were evaluated by determining the intraclass correlation coefficients (ICCs). To calculate the intraobserver ICCs, the MRI of 46 patients were randomly selected and segmented twice by radiologist A within one month. To calculate the interobserver ICCs, the selected images were independently segmented by two radiologists with 6 and 5 years of work experience in HCC imaging diagnosis. The intraobserver and interobserver ICCs were then calculated. An ICC greater than 0.75 was indicative of good consistency. Features with ICCs > 0.75 were retained and entered into the next step of screening; features with ICCs < 0.75 were excluded for each sequence.

The features were standardized via the zero-mean normalization method, and least absolute shrinkage and selection operator (LASSO) dimensionality reduction was used to screen the optimal radiomics features. In the process of feature selection, the LASSO penalty parameter (λ) was determined by 10-fold cross-validation, following the one standard error rule.

Univariate and multivariate analyses were used to select independent clinical factors associated with TACE treatment response. Using logistic regression, the FS-T2WI, AP, and PVP models and the joint-radiomic (JR) model were constructed on the basis of the optimal features from each MRI sequence. The independent clinical factors and all of the optimal radiomic features were integrated to construct a clinical-radiomic model (RC model). The performance of each model was evaluated via the construction of receiver operating characteristic (ROC) curves.

Statistical analysis

Statistical analysis was performed with R software (R4.2.0; https://www.r-project.org). An independent samples t test was used to analyze the differences in indicators that conformed to a normal distribution between the training group and the validation group. The Mann-Whitney U test was used to analyze the differences in indicators that did not conform to a normal distribution between the two groups. The χ² test was used to compare categorical variables. A two-sided P < 0.05 was considered to indicate statistical significance. The SHAP package in Python was used for the SHAP algorithm of the model.

RESULTS

Clinical features

Among the 116 patients in this study, 104 were males and 12 were females. The age range was 26-82 years, with a mean of 58.5 years. There were 52 patients with solitary lesions and 64 patients with multiple lesions. Fifty-two patients had liver cirrhosis. The ROC curve was drawn with TACE treatment response as the outcome variable and a cutoff value of 8.62 cm according to the maximum Youden index (area under the ROC curve of 0.649; 95% confidence interval: 0.541-0.757; P < 0.01; Figure 2). Patients were divided into two groups according to the maximum tumor diameter according to the cutoff value: ≤ 8.62 cm and > 8.62 cm. Among the neutrophil-to-lymphocyte ratio and platelet-to-lymphocyte ratio patients, the median was used as the cutoff value to divide patients into two groups. There were no significant differences in the clinical features of the patients between the training and validation groups (Table 1). Upon evaluation of the treatment response, 72 patients were placed in the response group, and 44 patients were placed in the nonresponse group. Univariate analysis revealed significant differences in the maximum tumor diameter, the number of tumors, and the presence or absence of portal vein tumor thrombus between the response group and the nonresponse group. Multivariate analysis revealed that the maximum tumor diameter and the number of tumors were independent predictors of the response to TACE among HCC patients (Table 2).

Figure 2 Assessment of the consistency of magnetic resonance imaging radiomic feature extraction via the intraclass correlation coefficient. A: Inter-observer agreement assessment of fat-saturation T2-weighted imaging; B: Intra-observer reliability assessment of fat-saturation T2-weighted imaging; C: Inter-observer agreement assessment of arterial phase; D: Intra-observer reliability assessment of arterial phase; E: Inter-observer agreement assessment of portal venous phase; F: Intra-observer reliability assessment of portal venous phase. ICC: Intraclass correlation coefficient.

Table 1 Clinical features of the patients.

Clinical parameter	Training group (n = 80)				Validation group (n = 36)				P value
	n	Response group (n = 50)	Nonresponse group (n = 30)	P value	n	Response group (n = 22)	Nonresponse group (n = 14)	P value
Sex				0.609				0.956	0.309
Male	73	45	28		31	19	12
Female	7	5	2		5	3	2
Age				0.954				0.204	0.908
≤ 60	45	28	17		21	11	10
> 60	35	22	13		15	11	4
Number of tumors				0.005				0.221	0.502
≤ 1	32	26	6		20	14	6
> 1	48	24	24		16	8	8
Maximum tumor diameter				0.003				0.005	0.212
≤ 8.62	44	34	10		23	18	5
> 8.62	36	16	20		13	4	9
ALBI				0.729				0.878	0.18
1	38	26	12		20	12	8
2	41	23	18		16	10	6
3	1	1	0		0	0
CTP				0.248				0.927	0.257
A	64	42	22		28	17	11
B	16	8	8		8	5	3
Liver cirrhosis				0.383				0.342	0.074
Yes	35	30	15		17	13	6
No	45	20	15		19	9	8
Portal vein tumor thrombus				0.023				0.311	0.629
Yes	14	45	21		14	12	10
No	66	5	9		22	10	4
NLR				0.862				0.4	1
≤ 3.375	39	24	15		16	11	5
< 3.375	41	26	15		20	11	9
PLR				0.602				0.311	0.4
≤ 114.385	37	22	15		22	12	10
> 114.385	43	28	15		14	10	4
ALT (U/L)				0.129				0.592	0.517
≤ 50	46	32	14		25	16	9
> 50	34	18	16		11	6	5
ALB (g/L)				0.729				0.878	1
≤ 40	42	27	15		20	12	8
> 40	38	23	15		16	10	6
AST (U/L)				0.313				0.927	0.571
≤ 40	24	17	7		8	5	3
> 40	56	33	23		28	17	11
AFP (ng/mL)				0.948				0.629	0.93
< 20	28	18	10		9	6	3
20-400	25	15	10		10	7	3
> 400	27	17	10		17	9	8

ALBI: Albumin-bilirubin; CTP: Child-Turcotte-Pugh; NLR: Neutrophil-to-lymphocyte ratio; PLR: Platelet-to-lymphocyte ratio; ALT: Alanine aminotransferase; ALB: Albumin; AST: Aspartate aminotransferase; AFP: Alpha-fetoprotein.

Table 2 Univariate and multivariate logistic regression analyses of the clinical features associated with transarterial chemoembolization treatment response.

Clinical feature	Univariate logistic regression analysis		Multivariate logistic regression analysis
	P value	OR (95%CI)	P value	OR (95%CI)
Sex	0.612	0.643 (0.117-3.541)
Age	0.954	0.973 (0.391-2.425)
Number of tumors	0.006	4.333 (1.512-12.416)	0.001	8.067 (2.276-28.594)
Maximum tumor diameter	0.003	4.250 (1.621-11.144)	0.001	7.637 (2.358-24.733)
Liver cirrhosis	0.384	1.500 (0.602-3.735)
Portal vein tumor thrombus	0.029	0.259 (0.077-0.869)
ALBI	0.409	1.443 (0.604-3.443)
CTP	0.252	1.909 (0.631-5.778)
NLR	0.862	0.923 (0.373-2.283)
PLR	0.603	0.786 (0.317-1.948)
ALT	0.131	2.032 (0.890-5.103)
ALB	0.729	1.174 (0.474-2.906)
AST	0.316	1.693 (0.605-4.735)
AFP	0.917	1.030 (0.596-1.777)

ALBI: Albumin-bilirubin; CTP: Child-Turcotte-Pugh; NLR: Neutrophil-to-lymphocyte ratio; PLR: Platelet-to-lymphocyte ratio; ALT: Alanine aminotransferase; ALB: Albumin; AST: Aspartate aminotransferase; AFP: Alpha-fetoprotein; OR: Odds ratio; CI: Confidence interval.

Consistency evaluation and feature dimensionality reduction

A total of 1223 radiomic features were extracted from each of the FS-T2WI, AP, and PVP datasets (Figure 2). After performing an independent samples t test or the Mann-Whitney U test, 1097, 1014 and 1055 significant features (P < 0.05) were screened. Finally, LASSO regression was used to select 3, select 2, and select 1 optimal feature from these features (Figure 3; Table 3).

Figure 3 Feature selection via least absolute shrinkage and selection operator regression. A: Least absolute shrinkage and selection operator regression coefficient path diagram. As the penalty parameter increases, the feature coefficients gradually decrease toward zero. The features with nonzero coefficients at the λ (1-standard error) line were ultimately selected, resulting in 3, 2, and 1 optimal feature from the features of the fat-saturation T2-weighted imaging, arterial phase, and portal venous phase datasets for subsequent model construction; B: Least absolute shrinkage and selection operator regression parameter diagram. The two vertical dashed lines indicate the selected values using cross-validation: The optimal value was obtained by applying the minimum criteria and 1 of the minimum criteria (1-standard error criteria).

Table 3 Radiomic features for predicting transarterial chemoembolization treatment response in hepatocellular carcinoma patients by least absolute shrinkage and selection operator regression analysis.

Cohort	Feature type	Feature name
FS-T2WI	Shape feature	Elongation
	GLSZM	Low-gray level zone emphasis
	GLDM	Gray level variance
Arterial phase	GLCM	Inverse variance
	First-order feature	Mean
Portal vein phase	First-order feature	Mean

FS-T2WI: Fat-saturation T2-weighted imaging; GLSZM: Gray level size zone matrix; GLDM: Gray level dependence matrix; GLCM: Gray level co-occurrence matrix.

Model evaluation

The area under the curve (AUC) values of the FS-T2WI, AP, PVP, JR model and RC model were 0.771, 0.668, 0.725, 0.845 and 0.929, respectively, in the training group and 0.705, 0.666, 0.675, 0.799 and 0.815, respectively, in the validation group. The clinical-radiomic model had the best predictive performance (Figure 4; Table 4).

Figure 4 Receiver operating characteristic curves. The showing the performance of the fat-saturation T2-weighted imaging, arterial phase, portal venous phase, joint-radiomic model, and radiomic-clinical model in predicting responses to transarterial chemoembolization. A: Training group; B: Validation group. FS-T2WI: Fat-saturation T2-weighted imaging; AP: Arterial phase; PVP: Portal venous phase; JR: Joint-radiomic; RC: Radiomic-clinical.

Table 4 Predictive performance of each model.

Cohort	Model	AUC	Sensitivity	Specificity	PPV	NPV	ACC	F1-score
Training	FS-T2WI	0.771	0.367	0.880	0.647	0.698	0.687	0.468
	AP	0.668	0.300	0.940	0.750	0.691	0.700	0.429
	PVP	0.725	0.233	0.940	0.700	0.671	0.675	0.350
	JR	0.845	0.600	0.940	0.857	0.797	0.813	0.706
	RC	0.929	0.833	0.920	0.862	0.902	0.888	0.847
Validation	FS-T2WI	0.705	0.429	0.772	0.545	0.680	0.639	0.480
	AP	0.666	0.357	0.864	0.625	0.679	0.667	0.455
	PVP	0.675	0.286	0.955	0.800	0.677	0.694	0.421
	JR	0.799	0.643	0.864	0.750	0.792	0.778	0.692
	RC	0.815	0.500	0.909	0.778	0.741	0.750	0.609

FS-T2WI: Fat-saturation T2-weighted imaging; AP: Arterial phase; PVP: Portal venous phase; JR: Joint-radiomic; RC: Radiomic-clinical; AUC: Area under the curve; PPV: Positive predictive value; NPV: Negative predictive value; ACC: Accuracy.

Interpretation analysis of the RC model

To further analyze the importance of the features included in the RC model, the SHAP algorithm was used to quantify the contributions of these features. The average absolute SHAP value of each feature across all the samples was calculated (Figure 5). The results revealed that the number of tumors had the greatest effect on the final prediction result in the model (Figure 5A and B), followed by the maximum tumor diameter. Additionally, we demonstrated how the model can be used to evaluate the treatment effect for two randomly selected samples with different outcomes (Figure 5C and D).

Figure 5 SHapley Additive exPlanations analysis plot of each predictor in the radiomic-clinical model. A: SHapley Additive exPlanations (SHAP) summary plot shows the distribution of the SHAP values of each predictor in all the samples; B: The SHAP bar plot shows the importance of each predictor in predicting early responses to transarterial chemoembolization (TACE); C: SHAP waterfall plot showing the predicted values for individual samples with the predicted outcome of non-response to TACE; D: SHAP waterfall plot showing the predicted values for individual samples with the predicted outcome of early response to TACE. AP: Arterial phase; PVP: Portal venous phase; FS-T2WI: Fat-saturation T2-weighted imaging; SHAP: SHapley Additive exPlanations.

DISCUSSION

In this study, the potential value of radiomic features from FS-T2WI, AP, and PVP images combined with clinical risk factors for predicting TACE treatment response in unresectable HCC patients was evaluated. The results showed that each MRI radiomic model could better predict the response to HCC TACE treatment, and the RC model established by combining radiomic features and clinical features achieved the best predictive performance. SHAP analysis revealed that the number of tumors among the clinical features and the AP mean among the radiomic features contributed the most to the model, followed by the maximum tumor diameter, PVP-mean and FS-T2WI low-gray level zone emphasis.

In recent years, several scholars have investigated the potential of using radiomics models based on a single MRI T2WI sequence, CE-MRI, or T2WI combined with CE-MRI AP images to predict responses to TACE among HCC patients. Weng et al[29] retrospectively analyzed the MRI data of 123 HCC patients who underwent TACE and constructed a model to predict responses to TACE on the basis of preoperative T2WI; the AUCs for the training group and the validation group were 0.812 and 0.801, respectively. Kong et al[22] collected pretreatment T2WI of 99 patients with advanced HCC who underwent TACE, screened radiomic features, calculated radiomic scores (Rad-scores), and constructed a prediction model; the AUCs for the training and validation cohorts were 0.812 and 0.866, respectively. Moreover, when the Rad-score was integrated with relevant clinical indicators to construct a new prediction nomogram, the AUCs increased to 0.861 and 0.884, respectively. Zhao et al[30] investigated the ability of a CE-MRI radiomic model to predict the response to TACE among HCC patients by extracting radiomic features in the arterial, venous, and delayed phases for modeling. The results revealed that among all the radiomic models, the three-phase radiomic model exhibited the best performance, with AUCs of 0.838 and 0.833 for the training group and the validation group, respectively. Additionally, a joint model integrating CE three-phase radiomic scores and clinical risk factors showed excellent predictive ability, with AUCs of 0.878 and 0.833 for the training and validation cohorts, respectively. Kuang et al[31] retrospectively analyzed 153 HCC patients with tumor diameters < 5 cm and constructed radiomic models to predict tumor response on the basis of T2WI and DCE-MRI AP images before TACE. The performances of the T2WI and DCE-MRI AP models were almost equal, while the performance of the nomogram based on MRI and clinical risk factors was greater than that of the radiomics models, with AUCs of 0.83 and 0.81, respectively, for the training and validation groups.

In this study, T2WI and CE-MRI AP and PVP radiomic features were combined with clinical risk factors to establish a joint clinical-radiomic prediction model that achieved good predictive performance, with AUCs of 0.929 and 0.815, respectively, for the training and validation groups. Recently, Luo et al[26] investigated the use of multiparametric MRI and radiomics to predict disease progression in advanced unresectable HCC patients treated via TACE in combination with lenvatinib by extracting radiomic features from seven sequences, including T1-weighted imaging (T1WI), AP T1WI, PVP T1WI, delayed-phase T1WI, T2WI, DWI, and ADC. Multivariate analysis revealed that tumor number and AP intensity enhancement were independent risk factors for disease progression, and the inclusion of radiomic features improved the disease progression prediction accuracy of the clinical models. Future studies should further compare the performance of the multiparametric MRI radiomics model in predicting tumor response. Some studies have shown that the efficacy of TACE in HCC patients is closely related to the status (number or size) of the tumors[32-36]. This study also revealed that the number of tumors and the maximum tumor diameter strongly correlated with the TACE treatment response, and SHAP analysis revealed their significant contributions to the RC model. Furthermore, integrating these clinical factors with radiomic features could further improve the performance of the prediction model.

Additionally, three methods are currently used in radiomics analysis to segment target lesions: Manual, automatic, and semiautomatic. In this study, the most common method, manual segmentation, was used in reference to Polan et al[37]. Compared with two-dimensional tumor sections, three-dimensional tumor volume data can provide more complete morphological information[27,38]. Moreover, the layer-by-layer segmentation method was used in this study to segment the target lesion (volume of interest) to extract radiomic features and obtain more comprehensive information on tumor heterogeneity.

This study has several limitations: (1) This was a retrospective study with a relatively small sample size; thus, a certain degree of selection bias may exist, potentially leading to overfitting. Further studies with larger sample sizes are needed; (2) The data used in this study were from a single center. Additional patients from other hospitals could be included to form an external validation set, thereby further validating the performance of the model; and (3) The follow-up time of this study was short, and continued follow-up is needed to verify the performance of the model in predicting long-term efficacy.

CONCLUSION

The interpretable RC model could stratify HCC patients into responders and nonresponders to TACE and serve as a potential tool to identify more appropriate HCC patients for TACE, thus sparing patients from receiving ineffective and unnecessary treatments[39].