Deep learning-assisted tumor radiomic dynamics on MRI predict pathological complete response in HCC undergoing immune-based therapy followed by hepatectomy

Abstract

Background and Aims:

Pathological complete response (pCR) following conversion therapy for initially unresectable hepatocellular carcinoma (uHCC) remains challenging to predict preoperatively. This study developed and validated a model integrating clinicopathological and radiomic features of the tumor to predict pCR.

Methods:

In this multicenter retrospective study, temporal radiomics features were extracted from baseline, post-treatment, and delta (change) MRIs. Serum AFP response was calculated as log₁₀(preoperative AFP)/log₁₀(baseline AFP). Univariate analysis, collinearity assessment, LASSO, and random forest were employed to perform feature selection. Fourteen machine learning models were benchmarked, with performance evaluated by using comprehensive metrics AUC, NPV, PPV, sensitivity, specificity, calibration, and decision curve analysis.

Results:

The model was developed and validated in a training (n=78), an internal test (n=32), and an independent validation cohort (n=44). The delta radiomic model significantly outperformed both baseline (test AUC: 0.835 vs. 0.483, p<0.05; validation AUC: 0.783 vs. 0.434, p<0.05) and preoperative models (test AUC: 0.685, p<0.05; validation AUC: 0.506, p<0.05), demonstrating superior predictive performance and generalization capability in predicting lesion-level pCR. Notably, when predicting patient-level pCR, the radiomic model also showed robust discrimination, with AUCs of 0.819 in the test set and 0.781 in the validation set. The combined radiomics-AFP model achieved even higher AUCs of 0.920 (test) and 0.857 (validation) in predicting lesion-level pCR.

Conclusions:

Dynamic radiomic changes effectively predict pCR in uHCC after conversion therapy. Combining delta radiomics with AFP response significantly improves predictive performance, offering a non-invasive method for assessing pCR and potentially guiding personalized treatment decisions.

INTRODUCTION

Hepatocellular carcinoma (HCC) is the sixth most common malignancy and the third leading cause of cancer-related mortality globally.¹ For patients with early-stage disease, surgical resection offers a 5-year survival rate of 50%–70%.² However, over 70% of patients present with intermediate or advanced HCC at their initial diagnosis.³

Systemic therapies (eg, atezolizumab–bevacizumab, apatinib–camrelizumab), administered alone or combined with locoregional modalities such as transarterial chemoembolization (TACE) and hepatic arterial infusion chemotherapy (HAIC), have achieved high objective tumor response rates.^4–9 This enables a subset of initially unresectable HCC (uHCC) patients to undergo curative surgical resection following conversion therapy.^10–13

Approximately 10%–40% of patients who undergo conversion therapy followed by resection achieve a pathological complete response (pCR), showing markedly superior 2-year progression-free survival (PFS) compared with their non-pCR counterparts (59.3% vs. 38.3%, p=0.045).¹⁴ However, the definitive diagnosis of pCR relies exclusively on the histological assessment of tumor specimens, which is unavailable preoperatively. Furthermore, significant discrepancies often exist between radiological evaluation and pathological response,¹⁵ making it challenging for preoperative imaging to precisely predict pCR. Notably, studies suggest that for patients who have achieved a radiological complete response, surgical resection may not confer a significant benefit in survival outcomes compared with non-operative management.^14,16 Currently, there is no reliable preoperative method for predicting pCR.

Radiomics is an emerging technology that leverages high-throughput extraction of quantitative features, such as texture, morphology, and functional heterogeneity, from medical images in combination with artificial intelligence algorithms to analyze the biological behavior of tumors.^17,18 In contrast to traditional imaging evaluation methods (eg, mRECIST criteria) that rely on the single parameter of enhancement degree, radiomics can mine over a thousand dimensions of sub-visual features. This allows for a comprehensive quantification of the spatial heterogeneity of the tumor microenvironment and the dynamic evolution of treatment response. Previous studies have demonstrated the limitations of using the depth of imaging response based on arterial phase enhancement to predict pCR, as this approach neglects tumor heterogeneity.¹⁹ Conversely, a systemic treatment response model that combines magnetic resonance imaging (MRI) radiomic features from the arterial and delayed phases has been shown to significantly improve the accuracy of efficacy prediction and exhibit universality for response stratification across different medical centers.²⁰

Given this context, we hypothesized that radiomic features could predict the achievement of pCR before hepatectomy. A predictive model was developed by combining various radiomic features, such as dynamic contrast-enhanced patterns, apparent diffusion coefficient heterogeneity from diffusion-weighted imaging, and features from plain scan images. The study aims to examine the effectiveness of this radiomic model in predicting pCR and to explore the impact of incorporating tumor marker response on predictive accuracy.

METHODS

Patients and treatment

This retrospective cohort study enrolled consecutive patients with initially unresectable HCC who were administered first-line systemic therapy (with or without locoregional therapy) and subsequently underwent hepatectomy from 3 tertiary hospitals.

Eligible participants met the following criteria: (i) clinically or pathologically confirmed HCC; (ii) received first-line systemic therapy (anti-angiogenic agents plus immune checkpoint inhibitors) or concurrent systemic therapy and locoregional therapy (TACE/HAIC); (iii) had both baseline (within 14 d pre-treatment) and preoperative (within 14 d pre-resection) dynamic contrast-enhanced MRI scans available in DICOM format; and (iv) had complete standardized pathological reports documenting pCR status. Exclusion criteria included atypical imaging features precluding a definitive clinical diagnosis of HCC, non-HCC histology, prior liver transplantation, incomplete imaging data, or an ambiguous pCR classification.

Preoperative conversion therapy primarily consisted of targeted therapy combined with immunotherapy or integrated locoregional therapies. Targeted agents included anti-vascular endothelial growth factor drugs, such as bevacizumab, and tyrosine kinase inhibitors (TKIs) such as lenvatinib, apatinib, and donafenib. Immunotherapy predominantly involved PD-1/PD-L1 inhibitors, including atezolizumab, sintilimab, pembrolizumab, tislelizumab, and camrelizumab.

A total of 110 patients from Zhongshan Hospital, Fudan University (April 2019 to May 2024) were randomly divided (7:3 ratio) into a training set and an internal test set. These patients were recruited from the prospective clinical trials: NCT04639284, NCT05389527, NCT04649489, and NCT04843943. An independent external validation cohort was then established, comprising 44 patients from Tianjin Medical University Cancer Institute & Hospital and Ruijin Hospital, Shanghai Jiao Tong University School of Medicine (January 2021 to August 2025), and supplemented by consecutive patients from Zhongshan Hospital, Fudan University, between July 2024 and July 2025. The study design is depicted in Figure 1.

FIGURE 1.

Flowchart of the study. Abbreviations: DWI, diffusion-weighted imaging; IP, in-phase T1-weighted imaging; MLP, multilayer perceptron; MRI, magnetic resonance imaging; OP, out-of-phase T1-weighted imaging; pCR, Pathological complete response; SVM, support vector machine; T1W, T1-weighted; T2W, T2-weighted.

This study was conducted in accordance with the principles of the Declaration of Helsinki and the Declaration of Istanbul. The research protocol was approved by the Zhongshan Hospital Research Ethics Committee (Approval Numbers: B2021-210). The requirement for informed consent was waived (retrospective computational analysis of images) by the Zhongshan Hospital Research Ethics Committee.

Clinical data collection

Demographic, laboratory, and survival data were extracted from electronic medical records. Baseline and preoperative liver function (Child–Pugh grade), as well as levels of tumor markers including alpha-fetoprotein (AFP) and protein induced by vitamin K absence or antagonist-II (PIVKA-II), were recorded. In addition, dates of disease recurrence and survival outcomes were documented.

Measurements

Prior to treatment, all patients underwent a baseline evaluation that included liver, renal, thyroid, adrenal, and cardiac function tests; a complete blood count; and testing for hepatitis B surface antigen (HBsAg), HBV DNA, and AFP.

The AFP response was defined as the ratio of the preoperative to baseline AFP logarithmic values, calculated as $\frac{\log 10\left(preoperativeAFP\right)}{\log 10\left(baselineAFP\right)}.$ ²¹ Similarly, the PIVKA-II response was calculated using the same formula for PIVKA-II values.

The delta radiomic value represents the change in radiomic features between baseline and preoperative scans, calculated as $\frac{{\mathit{radiomics}}_{\mathit{at}\mathit{baseline}}-{\mathit{radiomics}}_{\mathit{before}\mathit{surgical}\mathit{operation}}}{{\mathit{radiomics}}_{\mathit{at}\mathit{baseline}}}$ .

Recurrence-free survival (RFS) was defined as the time from surgery to disease recurrence or death. Overall survival (OS) was defined as the time from surgery to death from any cause. For event-free individuals, survival was censored at the last follow-up date.

Surgical and pathological assessment

Patients were deemed eligible for surgical resection if they met the following criteria: (1) feasibility of achieving R0 resection with sufficient remnant liver volume and function, (2) intrahepatic lesions showing a partial response (PR) or stable disease (SD) for at least 2 months, (3) absence of severe or persistent adverse effects from systemic therapy, and (4) no contraindications for hepatectomy.²² For these eligible patients, surgical resection was conducted according to standardized protocols, which included intraoperative ultrasonography for tumor localization and vascular mapping. Parenchymal transection was performed using ultrasonic dissection and clamp-crushing techniques, with inflow/outflow control via the Pringle maneuver or hepatic vein occlusion when necessary.²³

In our study, all pathological specimens were subjected to a central review by 2 dedicated pathologists to ensure consistency and accuracy in the assessment of pCR. The method of pCR assessment was documented in our previous study,¹⁹ which was aligned with the International Immunotherapy Pathological Criteria.²⁴ Briefly, a pCR was defined as the detection of no residual viable tumor cells on hematoxylin and eosin–stained slide sections among all resected primary tumor(s), tumor thrombosis, and metastatic lesions. Specifically, for multifocal tumors in this study, the 2 largest target lesions were selected for analysis.

Lesion-level and patient-level analysis strategy

Given the inclusion of patients with multifocal HCC, and heterogeneous responses existed in different tumor nodules, a hierarchical analysis strategy distinguishing between the lesion level and the patient level was employed in the present study.

As detailed in the pathological assessment section, for patients with multiple tumors, the 2 largest lesions per patient were designated as index lesions to enable rigorous one-to-one matching across pre-treatment imaging, post-treatment imaging, and histopathology. This ensured a reliable ground truth for model development. Furthermore, to ensure the integrity and comparability of paired data, any lesion identified in pre-treatment imaging that could not be definitively matched or located in post-treatment imaging was excluded from subsequent analysis. However, the final patient-level pathological complete response (pCR) status was determined by a comprehensive evaluation of all resected tumor tissue. The radiomics model was developed and trained at the lesion level. In the training cohort, this included 78 individual lesions from 78 patients. For validation in the test and external cohorts, a conservative patient-level prediction rule was adopted: a patient was classified as pCR only if all the evaluated target lesions were independently predicted as pCR by the lesion-level model.

MRI protocol

MRI was performed using a comprehensive protocol, as detailed in the Supplemental Methods, http://links.lww.com/HEP/K385. The acquisition included: T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), in-phase (IP) and out-of-phase (OP) T1-weighted imaging, as well as dynamic T1-weighted imaging in the pre-contrast phase (T1WI), late arterial phase (AP), portal venous phase (PVP), and delayed (for extracellular contrast agent) phase (DP).

Tumor segmentation and imaging preprocessing

The automatic segmentation of intrahepatic lesions was performed using 2 distinct approaches to ensure the robustness of the subsequent radiomics analysis. The primary delineations were generated by the artificial intelligence (AI) module of the uAI Research Portal.²⁵ To evaluate the reproducibility of the radiomics features derived from these automated segmentations, a second, independent set of segmentations was produced using a pre-trained nnU-Net-based deep learning algorithm applied across 8 MRI sequences. The intra-class correlation coefficient (ICC) was then calculated to assess the agreement between the feature sets extracted from the 2 methods. Only features demonstrating high reproducibility (ICC > 0.80) were retained for subsequent model development. Supplemental Figure S1, http://links.lww.com/HEP/K385, displays the ROI delineations for the same case at baseline and preoperatively. Radiomic features were extracted using the PyRadiomics package (version 3.1; https://pyradiomics.readthedocs.io) in Python 3.7, adhering to the Image Biomarker Standardisation Initiative (IBSI) guidelines (The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping).

A standardized preprocessing pipeline was applied to ensure multi-scanner data consistency. This involved spatial normalization via linear interpolation resampling to a 1.0 × 1.0 × 3.0 mm³ isotropic resolution, alongside N4 bias field correction to remove intensity inhomogeneities, thus establishing a reliable foundation for subsequent image fusion and quantitative analysis. Details of image preprocessing are presented in Supplemental Methods, http://links.lww.com/HEP/K385.

Feature extraction and feature selection

Prior to feature screening, all 8 MRI sequences were co-registered and integrated for analysis. A total of 2264 quantitative features were obtained from each sequence, categorized into 3 classes: 18 first-order statistics, 14 shape-based features, and 72 texture features. The texture features were further subdivided into 4 categories: Gray-Level Co-occurrence Matrix (GLCM) with 21 features, Gray-Level Dependence Matrix (GLDM) with 14 features, Gray-Level Run Length Matrix (GLRLM) with 16 features, and Gray-Level Size Zone Matrix (GLSZM) with 16 features.

In the training cohort, radiomic features with an ICC ≤ 0.8 were first excluded to ensure robustness and reproducibility, as previously reported. Subsequently, the Student t test was used to identify features with significant differences (p<0.05) between the pCR group and non-pCR groups. Spearman correlation analysis was then performed to evaluate multicollinearity; if the correlation coefficient between a pair of features was ≥0.9 or ≤−0.9, only the feature with superior predictive performance was retained. Furthermore, we used the least absolute shrinkage and selection operator (LASSO) logistic regression to further reduce the number of features. After integrating features and conducting preselection across all sequences, we identified the feature intersection using the following methods: (1) ranking the top 10 features based on random forest importance scores, (2) selecting the top 15 features from LASSO coefficient ranking, and (3) identifying the intersection features, which yielded the final features to develop the prediction model. This final set of features was then used to construct machine learning models.

Model development and assessment

The dynamic evolution of tumor heterogeneity during neoadjuvant therapy poses challenges for assessing treatment efficacy. To address this, the study defined 3 analytical time points: (1) at baseline (pre-treatment), (2) preoperative (post-treatment), and (3) delta radiomic features (capturing longitudinal changes of features between baseline and preoperative MRI).

Machine learning techniques were employed, with Z-score normalization applied for data standardization. Fourteen models were systematically evaluated, including logistic regression, tree-based algorithms, kernel-based support vector machine (SVM), and neural networks. Models were optimized via 5-fold cross-validation with a focus on maximizing sensitivity to identify potential pCR beneficiaries, while stipulating a pre-specified specificity target of >70%. The rationale stems from the greater clinical detriment of a false-positive result, which could lead to an unwarranted alteration of the surgical strategy, compared with a false-negative. A 70% threshold ensures a manageable risk for such false positives. This strategy of employing a high-specificity threshold is also prevalent in other literature.¹⁸

Model performance was evaluated using a multifaceted validation framework. Discrimination was quantified by the area under the receiver operating characteristic (ROC) curve, calculated using the bootstrap method (2000 resampling). Classification performance was assessed with a confusion matrix, from which accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were derived. Calibration was evaluated using a calibration curve, which assessed the reliability of class probability estimates. Clinical utility was determined by calculating the net clinical benefit using decision curve analysis (DCA) across a range of probability thresholds (10%–90%).

Statistical analysis

Categorical variables were expressed as counts and percentages and compared using the Pearson χ ² test, the Fisher exact test, or the Mann–Whitney U test, as appropriate. Continuous variables were expressed as mean±standard deviation or median (interquartile range, IQR) and were compared using the Student t test or the Mann–Whitney U test. Survival curves were generated using the Kaplan–Meier method. Variables with a p-value <0.05 in univariate logistic regression analyses were included in multivariate logistic regression analyses to identify independent factors of pCR. The DeLong test was employed to compare the AUCs between 2 models. A p-value of <0.05 was considered statistically significant. All statistical analyses were performed using Python (version 3.10.6).

RESULTS

Patient characteristics

The overall patient flow and cohort division are summarized in Figure 2. Patients were allocated into a training set (n=78), an internal test set (n=32), and an independent external validation set (n=44). The pCR rates were 32.1% (25/78) in the training cohort, 31.2% (10/32) in the test cohort, and 45.5% (20/44) in the validation cohort, with no significant difference observed across the groups (p=0.279) (Table 1).

FIGURE 2.

Patient inclusion and randomization flowchart of the study. Abbreviations: HCC, hepatocellular carcinoma; pCR, pathological complete response.

TABLE 1.

Baseline characteristics of patients across all cohorts

Variables	Training cohort (n=78)	Test cohort (n=32)	Validation cohort (n=44)	p
Age in year, mean±SD	56.6±10.1	60.3±11.2	59.3±11.1	0.186
Sex, n (%)				0.802
Female	8 (10.3)	3 (9.4)	6 (13.6)
Male	70 (89.7)	29 (90.6)	38 (86.4)
Locoregional therapy, n (%)				0.001
Yes	19 (24.4)	10 (31.2)	34 (77.3)
No	59 (75.6)	22 (68.8)	10 (22.7)
pCR, n (%)				0.279
Yes	25 (32.1)	10 (31.2)	20 (45.5)
No	53 (67.9)	22 (68.8)	24 (54.5)
ECOG PS, n (%)				0.419
0	71 (91.0)	29 (84.4)	41 (93.2)
1	7 (9.0)	5 (15.6)	3 (6.8)
Child–Pugh class, n (%)				0.106
A	77 (98.7)	32 (100.0)	41 (93.2)
B	1 (1.3)	0 (0.0)	3 (6.8)
HBsAg, n (%)				0.213
Negative	17 (21.8)	7 (21.9)	12 (27.3)
Positive	61 (78.2)	27 (78.1)	30 (68.2)
NA	0 (0.0)	0 (0.0)	2 (4.5)
HBV DNA, n (%)				0.127
<1000/mL	43 (55.1)	18 (56.2)	27 (61.4)
≥1000/mL	32 (41.0)	14 (43.8)	12 (27.3)
NA	3 (3.9)	0 (0.0)	5 (11.4)
AFP, ng/mL	299.0 (6.4, 7400.0)	452.0 (130.0, 16,311.0)	141.0 (7.8, 2409.5)	0.112
AFP, n (%)				0.696
<400 ng/mL	40 (51.3)	17 (53.1)	26 (59.1)
≥400 ng/mL	37 (47.4)	15 (46.9)	17 (38.6)
NA	1 (1.3)	0 (0.0)	1 (2.3)
PIVKA-II, mAU/mL	3280.5 (667.0, 21,269.0)	10,763.0 (424.5, 44,032.0)	1406.5 (295.0, 17,948.0)	0.279
PIVKA-II, n (%)				0.679
<1000 mAU/mL	23 (29.5)	10 (31.2)	18 (40.9)
≥1000 mAU/mL	53 (67.9)	21 (65.6)	24 (54.5)
NA	2 (2.6)	1 (3.1)	2 (4.5)
CNLC stage, n (%)				0.005
Ib	12 (15.4)	3 (9.4)	8 (18.2)
IIa	11 (14.1)	4 (12.5)	3 (6.8)
IIb	10 (12.8)	4 (12.5)	5 (11.4)
IIIa	32 (41.0)	20 (62.5)	17 (38.6)
IIIb	13 (16.7)	1 (3.1)	6 (13.6)
NA	0 (0.0)	0 (0.0)	5 (11.4)
BCLC stage, n (%)				0.011
A	12 (15.4)	3 (9.4)	8 (18.2)
B	21 (26.9)	8 (25.0)	8 (18.2)
C	45 (57.7)	21 (65.6)	23 (52.2)
NA	0 (0.0)	0 (0.0)	5 (11.4)
Targeted therapy, n (%)				0.001
Lenvatinib	37 (47.4)	11 (34.4)	22 (50.0)
Bevacizumab	38 (48.7)	18 (56.3)	14 (31.8)
Apatinib	2 (2.6)	3 (9.3)	5 (11.4)
Donafenib	1 (1.3)	0 (0.0)	0 (0.0)
NA	0 (0.0)	0 (0.0)	3 (6.8)
Anti-PD-1 antibody used, n (%)				0.001
Atezolizumab	24 (30.8)	13 (40.6)	8 (18.2)
Camrelizumab	11 (14.1)	4 (12.5)	13 (29.5)
Pembrolizumab	14 (17.9)	4 (12.5)	5 (11.4)
Sintilimab	24 (30.8)	8 (25.0)	12 (27.3)
Tislelizumab	5 (6.4)	2 (6.2)	6 (13.0)
Toripalimab	0 (0.0)	1 (3.1)	0 (0.0)

Abbreviations: AFP, alpha-fetoprotein; BCLC stage, Barcelona Clinic Liver Cancer; CNLC stage, Chinese Liver Cancer Staging System; ECOG PS, Eastern Cooperative Oncology Group Performance Status; HBsAg, hepatitis B surface antigen; pCR, pathologic complete response; PIVKA-II, protein induced by vitamin K absence or antagonist-II.

With data cutoffs in November 2025, median overall survival was not reached in any cohort in Supplemental Figure S2, http://links.lww.com/HEP/K385. Median recurrence-free survival was 26.4 months (training; 95% CI: 20.6–50.6), not reached (test), and 27.0 months (validation; 95% CI: 14.9–NE). pCR was associated with significantly prolonged OS and RFS across all cohorts. In the training and test cohorts, non-pCR patients had significantly shorter median RFS (21.6 and 19.4 mo, respectively) compared with pCR patients (not reached; both p<0.05). In the validation cohort, although non-pCR patients also exhibited a shorter median RFS (p=0.078) (Figure 3).

FIGURE 3.

Survival analysis stratified by pCR status in different patient cohorts. (A, B) Difference in OS between pCR and non-pCR groups in 2 cohorts. (C, D) Difference in RFS between pCR and non-pCR groups in 2 cohorts. Abbreviations: OS, overall survival; pCR, pathologic complete response; RFS, relapse-free survival.

Tumor marker dynamic changes as an independent predictor of pCR

A comparative analysis of characteristics between the pCR and non-pCR groups for all cohorts is summarized in Supplemental Table S1, http://links.lww.com/HEP/K385. A significantly lower AFP response was observed in pCR patients across all cohorts (all p<0.05): 0.2 (IQR: 0.1–0.4) versus 0.7 (IQR: 0.5–0.9) in training, 0.2 (IQR: 0.2–0.4) versus 0.9 (IQR: 0.4–1.0) in test, and 0.4 (IQR: 0.2–0.6) versus 0.5 (IQR: 0.4–1.0) in validation cohorts.

In the training cohort, univariate analysis identified AFP response (OR: 0.11, 95% CI: 0.04–0.29, p<0.001), baseline AFP (OR: 1.64, 95% CI: 1.03–2.62, p=0.037), preoperative PIVKA-II (OR: 0.00, 95% CI: 0.00–0.15, p=0.043), and PIVKA response (OR: 0.20, 95% CI: 0.10–0.40, p<0.001) as significant factors of pCR (Table 2), with only AFP response remaining significant in multivariate analysis (OR: 0.21, 95% CI: 0.07–0.69, p=0.010).

TABLE 2.

Univariate and multivariate logistic regression results of the training cohort

Variables	OR (95% CI)	p	OR (95% CI)	p
AFP at baseline, ng/mL	1.64 (1.03–2.62)	0.037	1.30 (0.57–2.94)	0.535
AFP before surgical operation, ng/mL	0.00 (0.00–178.89)	0.168
AFP response	0.11 (0.04–0.29)	<0.001	0.21 (0.07–0.69)	0.010
PIVKA-II at baseline, mAU/mL	1.30 (0.82–2.05)	0.261
PIVKA-II before surgical operation, mAU/mL	0.00 (0.00–0.15)	0.044	0.00 (0.00–inf)	0.212
PIVKA-II response	0.20 (0.04–0.29)	<0.001	0.43 (0.18–1.04)	0.062

Bold p values indicate statistical significance at the p<0.05 level. Abbreviations: AFP, alpha-fetoprotein; PIVKA-II, protein induced by vitamin K absence or antagonist-II.

Radiomic features predicted pCR for per lesion

Across all cohorts, a total of 159 individual lesions were analyzed. This comprised 78 lesions in the training cohort (corresponding to 78 patients), 36 lesions in the internal test cohort (corresponding to 32 patients, with 4 patients having two evaluable lesions), and 46 lesions in the external validation cohort (corresponding to 44 patients, with 2 patients having two evaluable lesions).

To evaluate imaging predictors at distinct time points, we constructed 3 radiomic models (baseline, preoperative, and delta). Details of the feature selection process and the final features selected for the delta radiomic model are provided in Supplemental Figure S3, http://links.lww.com/HEP/K385. Based on a benchmark test in the training cohort (Supplemental Figure S4, http://links.lww.com/HEP/K385), the multilayer perceptron (MLP) algorithm was selected for modeling.

The delta-radiomics model demonstrated markedly superior predictive performance and generalization capability compared with both baseline and preoperative models (Supplemental Table S2, http://links.lww.com/HEP/K385, Figure S5, http://links.lww.com/HEP/K385). The baseline model showed limited predictive value with AUCs of 0.483 (95% CI: 0.395–0.572) in the test cohort and 0.434 (95% CI: 0.273–0.602) in the validation cohort. The preoperative model achieved modest AUCs of 0.685 (test: 95% CI: 0.447–0.907) and 0.506 (validation: 95% CI: 0.343–0.680). In contrast, the delta model yielded substantially higher AUCs of 0.879 (95% CI: 0.794–0.950) in training, 0.835 (95% CI: 0.652–0.972) in testing, and 0.783 (95% CI: 0.644–0.900) in external validation. Furthermore, the performance of the delta-radiomics model was statistically superior to both the baseline and preoperative models across test and validation cohorts (all p<0.05). Besides, the delta model exhibited robust and well-balanced performance across all datasets, as evidenced by its accuracy, sensitivity, specificity, PPV, and NPV. Furthermore, a comparative analysis showed that the delta-radiomics model significantly outperformed the model based on mRECIST criteria, which yielded an AUC of 0.651 in the training (p<0.010), 0.700 in the test (p=0.256), and 0.610 (p=0.089) in the validation cohorts (Supplemental Figure S6, http://links.lww.com/HEP/K385).

Radiomic model-predicted pCR at the patient level

Given the presence of multiple lesions in both the test cohort and the validation cohort, we performed patient-level validation of pCR prediction for this radiomic model. As shown in Table 3, the model achieved an AUC of 0.819 (95% CI: 0.627–0.973) for predicting pCR at the patient level in the test cohort, with accuracy, PPV, NPV, sensitivity, and specificity of 0.794 (95% CI: 0.647–0.941), 0.615 (95% CI: 0.333–0.900), 0.905 (95% CI: 0.750–0.998), 0.800 (95% CI: 0.500–0.997), and 0.792 (95% CI: 0.619–0.957), respectively. In the independent validation cohort, the model also demonstrated robust discriminative ability: the AUC was 0.781 (95% CI: 0.632–0.908), accuracy was 0.750 (95% CI: 0.614–0.864), PPV was 0.680 (95% CI: 0.500–0.857), NPV was 0.844 (95% CI: 0.652–0.997), sensitivity was 0.848 (95% CI: 0.667–0.989), and specificity was 0.662 (95% CI: 0.464–0.857).

TABLE 3.

Predictive performance of different models for pCR at the patient level

	AUC	Acc	Sens	Spec	PPV	NPV
Clinicopathologic model
Training cohort	0.878 (0.767–0.954)	0.865 (0.790–0.938)	0.720 (0.536–0.885)	0.929 (0.857–0.984)	0.822 (0.643–0.966)	0.882 (0.796–0.953)
Test cohort	0.796 (0.635–0.933)	0.735 (0.588–0.882)	0.794 (0.500–1.000)	0.706 (0.500–0.885)	0.536 (0.273–0.778)	0.896 (0.737–1.000)
Validation cohort	0.754 (0.609–0.883)	0.717 (0.587–0.827)	0.501 (0.286–0.720)	0.919 (0.783–1.000)	0.842 (0.615–1.000)	0.667 (0.486–0.824)
Radiomic model
Training cohort	0.879 (0.794–0.950)	0.872 (0.795–0.936)	0.760 (0.600–0.913)	0.925 (0.849–0.982)	0.826 (0.667–0.960)	0.891 (0.807–0.964)
Test cohort	0.819 (0.627–0.973)	0.794 (0.647–0.941)	0.800 (0.500–0.997)	0.792 (0.619–0.957)	0.615 (0.333–0.900)	0.905 (95% CI: 0.750–0.998)
Validation cohort	0.781 (95% CI: 0.632–0.908)	0.750 (95% CI: 0.614–0.864)	0.848 (95% CI: 0.667–0.989)	0.662 (95% CI: 0.464–0.857)	0.680 (95% CI: 0.500–0.857)	0.844 (95% CI: 0.652–0.997)
Clinicopathologic–radiomic model
Training cohort	0.958 (0.906–0.992)	0.889 (0.815–0.951)	0.803 (0.636–0.950)	0.927 (0.855–0.983)	0.834 (0.678–0.963)	0.913 (0.833–0.980)
Test cohort	0.920 (0.802–1.000)	0.883 (0.765–0.971)	0.898 (0.667–1.000)	0.877 (0.727–1.000)	0.747 (0.462–1.000)	0.955 (0.850–1.000)
Validation cohort	0.857 (0.733–0.958)	0.720 (0.587–0.848)	0.819 (0.650–0.960)	0.623 (0.421–0.815)	0.667 (0.500–0.834)	0.791 (0.591–0.950)

Note: Data in parentheses are the 95% confidence intervals.

Abbreviations: Acc, accuracy; NPV, negative predictive value; PPV, positive predictive value; Sens, sensitivity; Spec, specificity.

Prediction of pCR by a combined clinicopathologic and radiomic model at the patient level

To further enhance predictive accuracy, we integrated the dynamic changes of serum biomarkers (AFP response) with delta radiomic features to construct a bimodal clinicopathologic–radiomic model.

As summarized in Table 3 and visualized in Figures 4A, B, the combined radiomics–clinical model exhibited strong predictive performance across both test and validation cohorts. In the test cohort, it achieved an AUC of 0.921, an accuracy of 0.882, a specificity of 0.875, and an NPV of 0.955. Corresponding values in the validation cohort were an AUC of 0.845, accuracy of 0.702, specificity of 0.600, and NPV of 0.789. Confusion matrix analysis further confirmed the model’s robustness, with 18 true positives and 15 true negatives in the validation set (p<0.001), and 9 true positives and 21 true negatives in the test set (p<0.001) (Figures 4C, D). Calibration curves demonstrated good agreement between predicted probabilities and observed outcomes (Supplemental Figure S7, http://links.lww.com/HEP/K385), while decision curve analysis showed a consistently higher net benefit of the model over the default strategies across a wide range of threshold probabilities (Figures 4E, F).

FIGURE 4.

Performance evaluation of the integrated clinical–radiomics model. (A, B) AUC comparison of 3 models: clinical-only, radiomics-only, and combined clinical–radiomics models in test and validation cohorts. (C, D) Confusion matrices of the integrated clinical–radiomics model in test and validation cohorts. (E, F) DCA is demonstrating the clinical utility of the integrated model in the test and validation cohorts. Abbreviations: DCA, decision curve analysis.

Global and local interpretation of the combined model

A SHapley Additive exPlanations (SHAP) analysis was used to interpret the clinical–radiomic model. The feature importance plot revealed that both AFP response and radiomic score were the most significant predictors of pCR (Figures 5A, B). Contribution plots demonstrated that higher AFP response values exhibited a negative correlation with the likelihood of pCR, whereas higher radiomic scores were positively correlated with pCR (Figures 5C, D).

FIGURE 5.

The SHAP analysis for the combination model. Abbreviations: AFP, alpha-fetoprotein; SHAP, SHapley Additive exPlanations.

DISCUSSION

In this study, we developed and validated a temporal radiomic framework that integrates baseline, preoperative, and delta radiomic features to predict pCR in patients with HCC. Our findings show that the delta-radiomics approach, which captured longitudinal changes during therapy, demonstrated superior predictive performance over baseline or preoperative models. Furthermore, this radiomics-based model significantly outperformed conventional mRECIST criteria. The most significant finding was that a bimodal integration strategy, combining dynamic radiomic features with a dynamic serological biomarker (AFP response), achieved the highest predictive accuracy, surpassing unimodal approaches. To our knowledge, this is the first study to successfully implement and validate a fused radiomics–clinical model for pCR prediction in HCC, addressing a critical unmet need in the assessment of response to conversion therapy.

Conventional MRI and even advanced techniques like diffusion-weighted imaging or PET/CT remain suboptimal for accurately identifying pCR in HCC.²⁶ This challenge is further compounded by the fact that at initial diagnosis, ~41%–75% of HCC patients present with multifocal tumors.²⁷ Multifocal HCC may arise from intrahepatic metastasis (IM) of a primary tumor or from multicentric occurrence (MO) of independent clones, leading to potential discrepancies between lesion-level and patient-level treatment responses.^28,29 Consequently, accurately assessing pCR at the patient level is clinically essential. Our delta-radiomics model, which captures dynamic changes during treatment, achieved AUCs of 0.835 in the test cohort and 0.756 in the validation cohort in predicting pCR at the lesion level, consistently outperforming static radiomic models (baseline AUC: 0.483 test/0.434 validation; preoperative AUC: 0.685 test/0.506 validation) and mRECIST criteria (AUC: 0.700 test/0.610 validation). Importantly, when applied to predict pCR at the patient level—addressing the heterogeneity in multifocal disease—the model also demonstrated robust performance, with AUCs of 0.819 in the test set and 0.781 in the validation set. This aligns with evidence from other cancers³⁰ and confirms that monitoring temporal radiomic evolution offers a more sensitive and comprehensive measure of treatment effect in HCC, both at the lesion and patient levels.

Beyond imaging phenotypes, dynamic serum AFP response emerged as a strong predictor of pCR, with significantly greater reductions in pCR patients (p<0.05). The integration of AFP response with delta radiomics created a powerful bimodal predictor, achieving an AUC of 0.921 in the test and 0.845 in the validation cohort. This combined model synergizes structural and microenvironmental information from imaging with biological tumor burden from serology, providing a robust tool for early treatment response assessment and personalized strategy optimization in HCC undergoing conversion therapy.^17,31,32

The strong predictive performance of our model suggests its potential to support clinical decision-making. Primarily, this model could serve as a valuable non-invasive tool to inform and enrich the preoperative dialogue between surgeons and patients. By providing a quantified probability of pCR, it helps to identify a subset of patients with an exceptionally favorable response to conversion therapy. In the future, for highly selected patients—such as those with high surgical risk or those who strongly desire organ preservation—a model-predicted high likelihood of pCR could support the exploration of a “watch-and-wait” strategy or less extensive surgical interventions. However, it is imperative to emphatically state that any deviation from the current standard of care (surgical resection) must be approached with utmost caution and should only be implemented within the context of a clinical trial. Our model is not intended to immediately replace surgery but to provide a robust, data-driven foundation for designing such future trials. The ultimate clinical utility would be to use the model to stratify patients, where those with a high predicted probability of pCR could be enrolled in trials investigating treatment de-escalation, while those with a low probability could be prioritized for aggressive resection or considered for alternative adjuvant therapies.

The 3 radiomic features ultimately incorporated into our model demonstrate plausible biological and clinical relevance. The feature wavelet_firstorder_wavelet-HLH-Skewness quantifies intensity distribution skewness after HLH wavelet transformation, a potential marker of intratumoral heterogeneity. Its analogues in pancreatic and breast cancers are associated with genomic alterations and proliferation indices (eg, Ki-67), supporting its potential as a non-invasive biomarker for tumor biology in HCC.³¹ The feature original_shape_Sphericity (quantifying lesion roundness) may predict pCR by capturing both macroscopic and biological therapy responses. An increase in sphericity reflects uniform necrosis and fibrotic shrinkage, while its association with the PI3K/AKT pathway³³ suggests concomitant inhibition of this pro-survival signaling, thereby collectively enabling tumor eradication. Lastly, original_shape_Maximum2DDiameterSlice, indicating the largest axial diameter within a single slice, provides a straightforward yet impactful measure of tumor burden. Its reduction post-therapy is a direct marker of tumor response,³⁴ and its integration into a radiomic model allows for a more nuanced assessment of dimensional changes. Collectively, these features provide a multifaceted characterization of the tumor’s phenotypic response to therapy, enhancing the biological interpretability of our predictive model.

Notwithstanding these promising findings, several limitations merit consideration. First, the retrospective design introduces a potential for selection bias. Second, while the variety in treatment regimens is a consideration, our analysis showed no evidence that it confounded the primary outcome (pCR) in this study. But further validation in larger populations is still necessary. Third, the relatively small sample size of the external validation cohort may limit the stability of performance estimates. Finally, our model currently relies solely on MRI-derived radiomic features. Future studies could enhance predictive power and biological interpretability by incorporating multi-parametric data, such as digital pathology whole-slide imaging and contrast-enhanced ultrasound, enabling histopathologic–spatial correlation analysis and offering a more comprehensive reflection of tumor heterogeneity.

In conclusion, we have developed and validated a predictive model that integrates dynamic clinical data and radiomic features from MRI to enable non-invasive and personalized assessment of pCR in initially uHCC patients following conversion therapy. This model holds promise as a practical and user-friendly tool to guide personalized treatment strategies for HCC patients undergoing conversion therapy, thereby enhancing clinical decision-making without the need for invasive procedures.

Supplementary Material

hep-84-453-s001.pdf ^{(1.1MB, pdf)}