Development and validation of a multivariable CT-based Delta-radiomics model for predicting efficacy to bevacizumab therapy in patients with colorectal liver metastases

Introduction

Colorectal cancer (CRC) is the third most common malignancy worldwide and ranks second in cancer-related mortality (1). Approximately 20–25% of patients with CRC develop colorectal liver metastases (CRLM) either at diagnosis or during follow-up (2). Although surgical resection remains the optimal strategy for improving overall survival (OS), fewer than 20% of patients with CRLM are eligible for curative resection (3,4). In recent years, bevacizumab, an antiangiogenic drug, has been recommended as a first-line regimen for conversion therapy in CRLM (5,6), significantly improving resection rates and patient survival (7). Nevertheless, patient responses to bevacizumab exhibit considerable heterogeneity; a subset of patients derive no benefit and may even experience disease progression (8,9). Consequently, the early and accurate assessment of bevacizumab efficacy is paramount for optimizing response rates and mitigating adverse events associated with ineffective treatments.

The Response Evaluation Criteria in Solid Tumors (RECIST) 1.1 serves as the established standard for evaluating treatment efficacy in CRLM (10). It provides a straightforward and objective framework based on morphological changes in tumor size observed on serial imaging before and after therapy (11). However, this criterion has inherent limitations, as it often fails to detect early microstructural alterations—such as changes in perfusion or necrosis—that precede measurable size reduction (12).

Radiomics has emerged as a promising non-invasive approach that extracts and quantitatively analyzes high-dimensional features from conventional medical images to uncover underlying tumor biology at a microscopic level (13). Previous studies (14-16) have demonstrated that preoperative computed tomography (CT)-based radiomics models can facilitate early prediction of chemotherapy response and progression-free survival in patients with CRLM. However, such models are inherently limited by their reliance on single-time-point imaging data and thus fail to capture the heterogeneous, dynamic alterations induced by tumor microenvironment remodeling during therapeutic intervention. In contrast, Delta-radiomics, as an emerging analytical paradigm, quantifies the changes in texture features before and after treatment, enabling more sensitive capture of the tumor’s biological response to treatment (17,18). This capacity to stratify patients according to temporal texture feature evolution holds substantial clinical promise for personalizing systemic therapy regimens and refining surgical conversion strategies (19). Recent evidence (20,21) indicates that CT-based delta-radiomics models outperform conventional static radiomics models in predicting first-line chemotherapy efficacy for patients with CRLM. Nevertheless, existing investigations have predominantly focused on first-line chemotherapy. To date, no study has systematically evaluated the utility of delta-radiomics for predicting treatment efficacy in CRLM patients receiving bevacizumab-combined chemotherapy.

Therefore, this study aimed to develop and validate a multivariable CT-based Delta-radiomics model to predict treatment efficacy in patients with CRLM receiving bevacizumab-combined chemotherapy. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0614/rc).

Methods

Patients

This multicenter retrospective study enrolled consecutive patients with CRLM who received bevacizumab treatment at three institutions between January 2018 and January 2023. Data from center 1 (Lanzhou University Second Hospital) were randomly divided into training cohorts and internal validation cohorts at a ratio of 7:3. The external validation cohorts were composed of the data from the other two centers (Sichuan Provincial People’s Hospital and Affiliated Hospital of Qinghai University). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional review boards of Lanzhou University Second Hospital (No. 2023A-379), Sichuan Provincial People’s Hospital (No. 2022-254) and Affiliated Hospital of Qinghai University (No. P-SL-202219), and individual consent for this retrospective analysis was waived.

The inclusion criteria were as follows: (I) histopathologically confirmed colorectal adenocarcinoma with at least one liver metastasis confirmed by biopsy; (II) treatment with bevacizumab combined with a standard first-line chemotherapy regimen [FOLFOX (folinic acid + fluorouracil + oxaliplatin), FOLFIRI (folinic acid + fluorouracil + irinotecan), CapeOX (capecitabine + oxaliplatin), or mXELIRI (capecitabine + irinotecan)]; and (III) availability of abdominal contrast-enhanced CT images at baseline, after 2 months (4 cycles), and 6 months (12 cycles). The exclusion criteria were as follows: (I) incomplete clinical or pathological data; (II) poor contrast-enhanced CT image quality due to patient respiration or motion artifacts; (III) liver metastases with a maximum diameter <1 cm; (IV) combination with other treatments before and during therapy, including arterial chemoembolization and radiofrequency ablation; (V) lack of follow-up data. The recruitment flow-chart of patients and liver metastases is shown in Figure 1.

Figure 1 The recruitment flow-chart of patients and liver metastases. Center 1, Lanzhou University Second Hospital; Center 2, Sichuan Provincial People’s Hospital; Center 3, Affiliated Hospital of Qinghai University. CRLM, colorectal liver metastases; CT, computed tomography; LM, liver metastasis.

CT acquisition

All CT scans were acquired using different multi-slice scanners (GE Discovery CT750 HD, GE Revolution CT, Philips ICT 128, or United Imaging uCT960+) across three participating centers. Patients fasted for ≥4 hours before scanning. During the scan, subjects were positioned supine with arms elevated above the head and received standardized breathing instructions. The scan range extended from the diaphragm to the inferior border of the pubic symphysis. The contrast agent (iodixanol, 320 mg/mL) was injected via an antecubital vein using a high-pressure injector at a flow rate of 3.5–4.0 mL/s and a dose of 1.0 mL/kg body weight. Detailed abdominal contrast-enhanced CT parameters for all scanners are summarized in Table 1.

Efficacy assessment

Following RECIST 1.1, a radiologist independently assessed the therapeutic response of liver metastases and patient efficacy based on CT images at baseline and 6 months post-treatment initiation. (I) Therapeutic response of liver metastases: the maximum diameter of liver metastases was measured at baseline and at 6 months to determine the therapeutic response. Liver metastases classified as complete response (CR) or partial response (PR) were deemed responsive, while those classified as stable disease (SD) or progressive disease (PD) were considered non-responsive. (II) Patient efficacy: given that most patients presented with multiple liver metastases, which might exhibit heterogeneous responses, a strict criterion was applied: the presence of PD in any liver metastases was classified as ineffective groups (non-responder), while others were classified as effective groups (responder).

Follow-up information was obtained through hospital medical records and regular telephone interviews, with a data cutoff date of February 2024. OS was defined as the time interval from the initiation of bevacizumab combined with chemotherapy to the date of death from any cause. Patients who were alive at the last follow-up were censored.

Clinical characteristics

Demographic and clinical data were extracted from the electronic medical record system. Variables included patient demographics (age, gender), tumor characteristics (primary location, type of liver metastasis, number of liver metastases, and maximum diameter of liver metastases), baseline serum tumor markers [carcinoembryonic antigen (CEA), cancer antigen 199 (CA199), and cancer antigen 125 (CA125)], histological grade, clinical T and N stage, molecular profiles (RAS and BRAF mutational status), and treatment regimens.

Image segmentation

The complete radiomics workflow is illustrated in Figure 2. Two radiologists, each with over 10 years of experience in abdominal imaging, performed volume of interest (VOI) segmentation using ITK-SNAP software. Segmentation was conducted on axial portal venous phase CT images acquired at baseline and 2 months post-treatment initiation. For radiomics analysis, up to 10 liver metastases were segmented, while carefully excluding adjacent liver parenchyma, vascular structures, and bile ducts. Radiomics features were subsequently extracted from these VOIs. To evaluate interobserver reproducibility, 78 liver metastases that were randomly selected from 25 patients underwent independent, blinded segmentation by both radiologists. Interobserver agreement was quantified using intraclass correlation coefficients (ICC), with features demonstrating an ICC ≥0.75 considered to have acceptable reproducibility and retained for further analysis.

Figure 2 The workflow for constructing and analyzing the predictive model of bevacizumab therapy efficacy in CRLM patients. Ratio refers to pre-treatment/post-treatment; Delta refers to (pre-treatment − post-treatment)/pre-treatment; DeltaABS refers to |(pre-treatment − post-treatment)/pre-treatment|. Center 1, Lanzhou University Second Hospital; Center 2, Sichuan Provincial People’s Hospital; Center 3, Affiliated Hospital of Qinghai University. CRLM, colorectal liver metastases; VOI, volume of interest.

Feature extraction and selection

All segmented volumes from the three participating institutions were resampled to an isotropic voxel size of 1×1×1 mm³ using linear interpolation. Image intensities were discretized into 25 gray levels using min-max normalization. Using Python, initial texture features were extracted from pre- and post-treatment images. Initially, 1,688 texture features were extracted from the images before and after the treatment respectively. Furthermore, temporal texture features were calculated to quantify temporal changes: (I) Ratio = pre-treatment/post-treatment; (II) Delta = (pre-treatment – post-treatment)/pre-treatment; (III) DeltaABS = |(pre-treatment – post-treatment)/pre-treatment|.

A multi-step feature selection strategy was employed to identify the most robust predictive features (Figure 2): (I) variance thresholding: features with zero or near-zero variance were removed, retaining 1,601 pre-treatment, 1,602 post-treatment, 1,499 Ratio, 1,554 Delta, and 1,554 DeltaABS features. (II) Boruta algorithm: feature importance was evaluated using the Boruta algorithm, reducing the pool to 6 pre-treatment, 7 post-treatment, 18 Ratio, 17 Delta, and 16 DeltaABS features. (III) Simulated annealing (SA): the SA algorithm was applied to optimize feature combinations, further narrowing the selection to 4 pre-treatment, 4 post-treatment, 7 Ratio, 4 Delta, and 3 DeltaABS features. (IV) Stepwise logistic regression: finally, multivariable stepwise logistic regression identified the optimal subset comprising 3 pre-treatment, 1 post-treatment, 2 Ratio, 1 Delta, and 2 DeltaABS features.

Model construction and evaluation

Eight logistic regression models were constructed in the training cohorts to predict therapeutic response of liver metastases (Figure 2): (I) clinical model: based on statistically significant clinical parameters identified in univariate analysis. (II) Radiomics model: including pre-treatment, post-treatment, combined pre- and post-treatment (Pre + Post), Ratio, Delta, and DeltaABS models, each utilizing its optimal feature subset. (III) COMB model: integrated nine optimal texture features.

The prediction performances of the models were evaluated using the receiver operating characteristic (ROC) curves and area under the curve (AUC) in both the internal and external validation cohorts. Diagnostic metrics, including accuracy (ACC), sensitivity (SEN), specificity (SPC), negative predictive value (NPV), and positive predictive value (PPV), were calculated based on optimal cutoff thresholds determined by Youden’s index. Statistical differences among AUCs were assessed using DeLong’s non-parametric test. Model calibration was assessed via calibration curves, and clinical utility was evaluated using decision curve analysis (DCA) to estimate net benefit across a range of probability thresholds.

Statistical analysis

All statistical analyses were performed using Python (version 3.7.1; https://www.python.org) and R software (version 4.1.3; https://www.rproject.org). The normality of quantitative variables was assessed using the Kolmogorov-Smirnov test. Normally distributed data were presented as mean ± standard deviation, while non-normally distributed data were expressed as median with interquartile range (IQR). Categorical variables were reported as frequencies and percentages. Differences in clinical features between groups were evaluated using the Student’s t-test, Mann-Whitney U test, or chi-square test, as appropriate. The ICC was used to evaluate the interobserver agreement, with the following criteria: poor (<0.50), moderate (0.50–0.74), good (0.75–0.89), and excellent (≥0.90). Kaplan-Meier curves and log-rank tests were used to assess OS. Univariable and multivariable Cox proportional hazard regression analyses were used to identify independent predictors of OS, with hazard ratios (HRs) and 95% confidence intervals (CIs) calculated. Statistical significance was set at P<0.05.

Results

Patients

A total of 90 patients with CRLM and 255 liver metastases treated with bevacizumab combined with chemotherapy were included in this study. The treatment regimens included FOLFOX (n=23, 25.56%), FOLFIRI (n=15, 16.67%), CapeOX (n=42, 46.67%), and mXELIRI (n=10, 11.11%). The patients were split into a training cohort (47 patients, 125 liver metastases), an internal validation cohort (9 patients, 53 liver metastases) and an external validation cohort (34 patients, 77 liver metastases). Following patient efficacy assessment, 30 (63.83%), 4 (44.44%), and 18 (52.94%) patients were identified as responders in the training, internal validation, and external validation cohorts, respectively. There was no statistically significant difference in basic characteristics between the combined training and internal validation cohorts and the external validation cohorts (P>0.05). The basic information of the patients is summarized in Table 2.

Univariate analysis of clinical parameters and patient efficacy

There was a statistically significant difference in the distribution of RAS and BRAF mutation statuses between the two groups (P<0.05, Table 3). Subgroup analysis revealed that patients carrying RAS mutations and BRAF wild-type exhibited a better therapeutic response. Furthermore, stratified analysis by therapeutic response of liver metastases consistently demonstrated that the differences in RAS and BRAF mutation statuses between the two groups remained statistically significant.

Feature selection and model performance

Feature extraction performed by the two physicians demonstrated good interobserver agreement, with ICC values exceeding 0.80. Following feature screening and dimensionality reduction, a total of nine optimal texture features were retained. Violin plots revealed statistically significant differences in these features between the two groups stratified by the therapeutic response of liver metastases (P<0.05, Figure 3).

Figure 3 Comparison of nine optimal texture features between the responsive (PR) and non-responsive group (PD/SD). PD, progressive disease; PR, partial response; SD, stable disease.

The clinical model yielded AUC values of 0.728, 0.761, and 0.514 in the training, internal validation, and external validation cohorts, respectively. Its predictive performance was inferior to that of most radiomics models across all three cohorts. Among the radiomics models constructed using initial texture features, the post-treatment model and Pre + Post model were superior to the pre-treatment model in the external validation cohorts, with AUC values of 0.746, 0.735 and 0.435 respectively. Furthermore, within the subset of models based on temporal texture features, the Ratio and Delta models demonstrated significantly superior predictive performance compared to the DeltaABS model in the external validation cohorts, with AUC values of 0.922, 0.859 and 0.488, respectively (Figure 4A).

Figure 4 (A) ROC, (B) calibration, and (C) decision curve of the model used to predict the therapeutic response of liver metastases. The left panel is the training cohort, the middle panel is the internal validation cohort, and the right panel is the external validation cohort. Pre refers to pre-treatment; Post refers to post-treatment; Pre + Post refers to before and after treatment; Ratio refers to pre-treatment/post-treatment; Delta refers to (pre-treatment − post-treatment)/pre-treatment; DeltaABS refers to |(pre-treatment − post-treatment)/pre-treatment|; COMB refers to the integration of nine optimal texture features. AUC, area under the curve; CI, confidence interval; HL, hosmer-lemeshow; ROC, receiver operating characteristic.

Compared with the aforementioned models, the COMB model, which integrated the nine optimal texture features, demonstrated superior predictive performance (Figure 4A and Table 4). The DeLong test showed statistically significant differences in AUC between the COMB model and the clinical, pre-treatment and DeltaABS models in the external validation cohorts (P<0.05, Z=4.938, −5.355, −4.667). In contrast, no significant differences were found when comparing the COMB model with the post-treatment, Pre + Post, Ratio, or Delta models (P>0.05, Z=−1.485, −1.623, 2.004, 0.685). Calibration curves and DCA demonstrated that the prediction probabilities of the Ratio, Delta, and COMB models were highly consistent with the observed results and had good clinical utility (Figure 4B,4C).

The radiomics score of the COMB model was calculated as follows: radiomics score = −1.595 × Pre_original_glcm_MCC + 0.652 × Pre_wavelet.HLH_glszm_GrayLevelNonUniformityNormalized + 0.718 × Pre_logarithm_firstorder_Entropy + 0.994 × Pos_log.sigma.5.0.mm.3D_firstorder_Median + 2.112 × Ratio_wavelet.LLL_glcm_InverseVariance − 3.091 × Ratio_lbp.3Dm2_gldm_DependenceEntropy + 4.573 × Delta_log.sigma.1.0.mm.3D_glszm_SizeZoneNonUniformity + 2.259 × ABS_original_glrlm_RunEntropy + 2.073 × ABS_wavelet.LLL_gldm_DependenceNonUniformityNormalized.

Univariate and multivariate associations of covariates with OS

Univariate Cox regression analysis revealed that CA125 levels (HR: 1.003; 95%CI: 1.001–1.006, P=0.01) and patient efficacy (HR: 3.017; 95% CI: 1.192–3.449; P<0.001) were significantly associated with OS in patients with CRLM. Subsequent multivariate Cox analysis demonstrated that only patient efficacy (HR: 2.842; 95% CI: 1.468–5.502; P=0.002) remained an independent prognostic factor for OS (Table 5). Furthermore, the Kaplan-Meier curves demonstrated that patient efficacy could effectively stratify patients with CRLM into distinct prognostic groups (Figure 5).

Figure 5 Kaplan-Meier survival analysis of patient efficacy. (A) Training cohorts and internal validation cohorts; (B) external validation cohorts. CI, confidence interval; HR, hazard ratio; OS, overall survival.

Discussion

Early and accurate identification of the efficacy of bevacizumab-combined chemotherapy in patients with CRLM is essential to guiding personalized treatment decisions and enabling surgical conversion strategies. In this multicenter study, we developed and validated eight models to predict the therapeutic response of liver metastases at 6 months, using clinical variables, baseline and 2-month post-treatment CT texture features. The results demonstrated that the models (Ratio, Delta, and COMB) using temporal texture feature metrics outperformed those relying solely on clinical variables or initial texture features, with AUC of 0.858–0.956 in training, 0.891–0.899 in internal validation, and 0.833–0.922 in external validation cohorts. Furthermore, the patient efficacy at 6 months was confirmed to be the sole independent predictor of OS, supporting the utility of early efficacy assessment for risk stratification in patients with CRLM.

The NCCN guidelines (22,23) recommend RAS and BRAF mutational status as a critical determinant for selecting anti-angiogenic therapies and evaluating treatment efficacy in patients with CRLM. While RAS/BRAF mutations are established drivers of abnormal MAPK pathway activation and subsequent vascular endothelial growth factor (VEGF) upregulation, their direct predictive value for anti-VEGF efficacy remains complex (24). The findings of this study indicate that patients harboring RAS mutations and with BRAF wild-type may benefit from anti-VEGF treatments, such as bevacizumab. A meta-analysis (25) confirmed significant concordance in RAS/BRAF status between primary colorectal tumors and paired liver metastases, with minimal heterogeneity across multiple metastases within individual patients. Consequently, we assigned the RAS/BRAF status of the primary tumor to all liver metastases in patients with CRLM for univariate analysis. Our findings showed that RAS-mutated liver metastases exhibited superior treatment responses compared to the wild-type, while BRAF-mutated metastases demonstrated inferior responses, consistent with previous research (26). We further developed a clinical model that integrated the RAS/BRAF data to predict the therapeutic response of liver metastases. Although this model showed some predictive capability (AUC >0.5), it underperformed relative to most radiomics models. This suggests that pre- and post-treatment CT texture features may represent more promising biomarkers for predicting bevacizumab efficacy than clinical parameters alone.

After feature extraction and selection, we identified nine optimal texture features, comprising two first-order and seven second-order features. Our findings confirmed that elevated entropy (reflecting tumor heterogeneity) and median values (indicating hypervascular characteristics in tumor hemodynamics) among first-order features were significantly associated with a poorer therapeutic response of liver metastases, which is consistent with previous studies (15,16). This phenomenon can be explained by the joint representation of a proangiogenic tumor phenotype, where increased entropy and median values signify abnormal proliferation of tumor microvasculature, forming a heterogeneously perfused microenvironment. Under anti-VEGF therapy, this microenvironment promotes compensatory activation of the Ang-2/Tie2 pathway, ultimately driving treatment resistance (27,28). In addition, the seven second-order features, such as gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, and gray-level dependence matrix, collectively characterize the lesion texture homogeneity and complexity. These metrics reflect microscopic variations in tumor enhancement patterns, internal architecture, and microenvironmental organization (27,29). While Liu et al. (30) have associated higher coefficients of second-order features with increased tumor heterogeneity and targeted therapy resistance, our data linked certain high-coefficient features to a better therapeutic response. This discrepancy may indicate that the temporal texture features have new meanings.

The value of temporal texture features in predicting therapeutic response of liver metastases was extensively analyzed. Consistent with prior studies (20,21), these features outperformed initial texture features. Notably, we refined the Delta feature algorithm by adopting a normalized calculation—(pre-treatment − post-treatment)/pre-treatment—departing from the absolute difference approach used by Giannini et al. (20). Meanwhile, we introduced a novel Ratio feature that captures relative change rates rather than absolute differences, significantly improving model performance and cross-center generalizability. This enhancement likely confers greater stability and broader applicability to Delta-radiomics frameworks. Furthermore, the COMB model, integrating nine optimal texture features, achieved superior predictive performance (AUC >0.8), underscoring the added value of multivariable feature fusion.

Numerous studies (31-33) have established the value of CT-texture-based radiomics models for predicting the therapeutic response and OS in patients with CRLM, with AUC exceeding 0.75. However, these models were derived exclusively from single first-line chemotherapy cohorts, limiting their applicability to bevacizumab-combined regimens. Among the eight models evaluated, our multicenter study ultimately developed and validated three multivariable Delta-radiomics models tailored to bevacizumab combined with chemotherapy, supporting adaptive treatment decision-making. In addition, we further demonstrated that the patient efficacy defined by RECIST 1.1 at 6 months was significantly associated with OS (P=0.002). Based on these findings, we propose a clinical decision-making pathway for patients with CRLM receiving bevacizumab treatment: “model prediction → efficacy assessment → survival estimation → therapeutic strategy refinement”. This pathway entails predicting the therapeutic response of liver metastases, followed by objective evaluation of patient efficacy and survival prognosis, culminating in evidence-informed refinement of the personalized treatment strategies. Critically, the pathway enables early stratification of responders versus non-responders as early as 2 months into systemic therapy, thereby facilitating timely multidisciplinary team intervention prior to radiographic or clinical progression. For predicted responders, continuation of the current regimen is supported by the model—with planned curative-intent resection. In contrast, for predicted non-responders, therapeutic escalation is recommended: either switching systemic therapy (e.g., from anti-VEGF to anti-EGFR agents) or integrating locoregional modalities (e.g., stereotactic body radiotherapy or radiofrequency ablation) (22,23). Such proactive adaptation helps avoid prolonged administration of ineffective bevacizumab-based regimens, preserves the window for surgical conversion, and mitigates unnecessary hepatic toxicity associated with overtreatment.

This study had some limitations. First, although designed as a multicenter retrospective analysis, the sample size remained relatively modest. Future investigations should expand cohort sizes to enhance the prediction performance. Second, manual segmentation of liver metastases introduces potential variability; implementing semi-automated or fully automated segmentation methods would improve efficiency and reproducibility. Pathological validation of the therapeutic response of liver metastases using tumor regression grade is warranted in future studies.

Conclusions

In conclusion, we developed and validated three multivariable CT-based Delta-radiomics models that utilize baseline and 2-month texture features to accurately predict 6-month treatment efficacy in patients with CRLM undergoing bevacizumab therapy. Importantly, the clinical decision-making pathways constructed based on these models can achieve early patient stratification, allowing clinicians to promptly identify non-responders and tailor personalized strategies.

Acknowledgments

The abstract of this study has been accepted for an oral presentation at the 2026 European Radiology Congress (#20068). We would like to thank Editage (www.editage.cn) for English language editing.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0614/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0614/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0614/prf

Funding: This work was supported by grants from the National Natural Science Foundation of China (No. 82371914), the Major Project of the Gansu Province Joint Research Fund (No. 24JRRA921), and the Cuiying Scientific and Technological Innovation Program of Lanzhou University Second Hospital (No. CY2021-ZD-01).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0614/coif). J.Z. obtained the project funds from the National Natural Science Foundation of China (No. 82371914), the Major Project of the Gansu Province Joint Research Fund (No. 24JRRA921), and the Cuiying Scientific and Technological Innovation Program of Lanzhou University Second Hospital (No. CY2021-ZD-01) to support the funding of this article and the article processing fee. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional review boards of Lanzhou University Second Hospital (No. 2023A-379), Sichuan Provincial People’s Hospital (No. 2022-254) and Affiliated Hospital of Qinghai University (No. P-SL-202219), and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Li M, Cao S, Xu RH. Global trends and epidemiological shifts in gastrointestinal cancers: insights from the past four decades. Cancer Commun (Lond) 2025;45:774-88. [Crossref] [PubMed]
2. Görgec B, Hansen IS, Kemmerich G, et al. MRI in addition to CT in patients scheduled for local therapy of colorectal liver metastases (CAMINO): an international, multicentre, prospective, diagnostic accuracy trial. Lancet Oncol 2024;25:137-46. [Crossref] [PubMed]
3. Li G, Zhu B, Li J, et al. Path forward: Cryoablation for colorectal liver metastases. Eur J Radiol 2026;199:112817. [Crossref] [PubMed]
4. Liu WX, Jiang C, Li JX. Current status and advances in the treatment of colorectal cancer with liver metastases. World J Clin Oncol 2025;16:107438. [Crossref] [PubMed]
5. Cohen R, Raeisi M, Chibaudel B, et al. Prognostic value of liver metastases in colorectal cancer treated by systemic therapy: An ARCAD pooled analysis. Eur J Cancer 2024;207:114160. [Crossref] [PubMed]
6. Bond MJG, Bolhuis K, Loosveld OJL, et al. First-Line Systemic Treatment for Initially Unresectable Colorectal Liver Metastases: Post Hoc Analysis of the CAIRO5 Randomized Clinical Trial. JAMA Oncol 2025;11:36-45. [Crossref] [PubMed]
7. Adam R, Piedvache C, Chiche L, et al. Liver transplantation plus chemotherapy versus chemotherapy alone in patients with permanently unresectable colorectal liver metastases (TransMet): results from a multicentre, open-label, prospective, randomised controlled trial. Lancet 2024;404:1107-18. [Crossref] [PubMed]
8. Varnier R, Toullec C, Philonenko S, et al. Triplet chemotherapy with or without bevacizumab as first line treatment for metastatic colorectal cancer: An AGEO multicenter real-world study. Dig Liver Dis 2024;56:1605-13. [Crossref] [PubMed]
9. Zheng X, Huang F, Zhan D, et al. Body composition analysis for predicting response to first-line bevacizumab in patients with unresectable colorectal liver metastases: A multicenter study. Eur J Radiol 2026; Epub ahead of print. [Crossref] [PubMed]
10. Eisenhauer EA, Therasse P, Bogaerts J, et al. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). Eur J Cancer 2009;45:228-47. [Crossref] [PubMed]
11. Michiel Zeeuw J, Kemna R, Ali M, et al. Total tumor volume predicts overall survival and response to induction chemotherapy in patients with colorectal cancer liver metastases: An ancillary study of the phase 3 CAIRO5 trial. Eur J Cancer 2025;228:115738. [Crossref] [PubMed]
12. Li S, Yuan L, Yue M, et al. Early evaluation of liver metastasis using spectral CT to predict outcome in patients with colorectal cancer treated with FOLFOXIRI and bevacizumab. Cancer Imaging 2023;23:30. [Crossref] [PubMed]
13. Wu J, Xia Y, Wang X, et al. Radiomics++: Review of Habitat Imaging Analysis for Decoding Tumor Heterogeneity. Annu Rev Biomed Eng 2026;28:219-48. [Crossref] [PubMed]
14. Kha QH, Nguyen PK, Le MHN, et al. A CT radiomics signature enables risk stratification and survival prediction in colorectal liver metastases. Phys Med Biol 2026;71: [Crossref] [PubMed]
15. Jing HH, Hao D, Liu XJ, et al. Development and validation of a radiopathomics model for predicting liver metastases of colorectal cancer. Eur Radiol 2025;35:3409-17. [Crossref] [PubMed]
16. Zhan G, Li Z, Liu X, et al. Combined dual-phase CT radiomics and deep transfer learning model for oxaliplatin response in unresectable colorectal liver metastases. Abdom Radiol (NY) 2026;51:3323-35. [Crossref] [PubMed]
17. Nardone V, Patanè V, Marinelli L, et al. Delta-Radiomics Biomarker in Colorectal Cancer Liver Metastases Treated with Cetuximab Plus Avelumab (CAVE Trial). Diagnostics (Basel) 2025;15:2914. [Crossref] [PubMed]
18. Nardone V, Reginelli A, Rubini D, et al. Delta radiomics: an updated systematic review. Radiol Med 2024;129:1197-214. [Crossref] [PubMed]
19. van der Velden AL, Knapen RRMM, Rahmani H, et al. The influence of AI on surgical and ablative treatments for colorectal liver metastases: a review of the current literature. Insights Imaging 2025;16:253. [Crossref] [PubMed]
20. Giannini V, Pusceddu L, Defeudis A, et al. Delta-Radiomics Predicts Response to First-Line Oxaliplatin-Based Chemotherapy in Colorectal Cancer Patients with Liver Metastases. Cancers (Basel) 2022;14:241. [Crossref] [PubMed]
21. Ye S, Han Y, Pan X, et al. Association of CT-Based Delta Radiomics Biomarker With Progression-Free Survival in Patients With Colorectal Liver Metastases Undergo Chemotherapy. Front Oncol 2022;12:843991. [Crossref] [PubMed]
22. Benson AB, Venook AP, Adam M, et al. Colon Cancer, Version 3.2024, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2024;22:e240029. [Crossref] [PubMed]
23. Benson AB, Venook AP, Adam M, et al. NCCN Guidelines® Insights: Rectal Cancer, Version 3.2024. J Natl Compr Canc Netw 2024;22:366-75. [Crossref] [PubMed]
24. Parseghian CM, Sun R, Woods M, et al. Resistance Mechanisms to Anti-Epidermal Growth Factor Receptor Therapy in RAS/RAF Wild-Type Colorectal Cancer Vary by Regimen and Line of Therapy. J Clin Oncol 2023;41:460-71. [Crossref] [PubMed]
25. Bhullar DS, Barriuso J, Mullamitha S, et al. Biomarker concordance between primary colorectal cancer and its metastases. EBioMedicine 2019;40:363-74. [Crossref] [PubMed]
26. Bond MJG, Bolhuis K, Loosveld OJL, et al. First-line systemic treatment strategies in patients with initially unresectable colorectal cancer liver metastases (CAIRO5): an open-label, multicentre, randomised, controlled, phase 3 study from the Dutch Colorectal Cancer Group. Lancet Oncol 2023;24:757-71. [Crossref] [PubMed]
27. van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
28. Frentzas S, Simoneau E, Bridgeman VL, et al. Vessel co-option mediates resistance to anti-angiogenic therapy in liver metastases. Nat Med 2016;22:1294-302. [Crossref] [PubMed]
29. Whybra P, Zwanenburg A, Andrearczyk V, et al. The Image Biomarker Standardization Initiative: Standardized Convolutional Filters for Reproducible Radiomics and Enhanced Clinical Insights. Radiology 2024;310:e231319. [Crossref] [PubMed]
30. Liu P, Zhu H, Zhu H, et al. Predicting Survival for Hepatic Arterial Infusion Chemotherapy of Unresectable Colorectal Liver Metastases: Radiomics Analysis of Pretreatment Computed Tomography. J Transl Int Med 2022;10:56-64. [Crossref] [PubMed]
31. Zhou S, Sun D, Mao W, et al. Deep radiomics-based fusion model for prediction of bevacizumab treatment response and outcome in patients with colorectal cancer liver metastases: a multicentre cohort study. EClinicalMedicine 2023;65:102271. [Crossref] [PubMed]
32. Qi W, Yang J, Zheng L, et al. CT-based radiomics for the identification of colorectal cancer liver metastases sensitive to first-line irinotecan-based chemotherapy. Med Phys 2023;50:2705-14. [Crossref] [PubMed]
33. Nakanishi R, Oki E, Hasuda H, et al. Radiomics Texture Analysis for the Identification of Colorectal Liver Metastases Sensitive to First-Line Oxaliplatin-Based Chemotherapy. Ann Surg Oncol 2021;28:2975-85. [Crossref] [PubMed]