Intratumoral and peritumoral delta radiomics of MRI predicts overall survival to targeted therapy in colorectal cancer with liver metastases

doi:10.21203/rs.3.rs-8185563/v1

Background: To develop and validate a hybrid prognostic model integrating clinical, MRI, and radiomics features for predicting overall survival (OS) in patients with colorectal cancer liver metastases (CRCLM) undergoing targeted therapy.

Methods: This retrospective, multicenter study included 118 CRCLM patients who received targeted therapy at two tertiary hospitals. Patients were divided into training, internal test, and external validation cohorts. Clinical, MRI, and radiomics features were comprehensively collected. Multiple radiomics models—including intratumoral, peritumoral, combined, and delta models—were constructed. The optimal model was selected and combined with key clinical and imaging features to build a hybrid prognostic model. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration, and risk stratification by Kaplan-Meier analysis.

Results: Focal extranodular protuberant (FEP) lesions, tumor related to adjacent vein (TRTAV), and the radiomics-derived Rad-score were identified as independent predictors of OS. The hybrid model demonstrated superior prognostic accuracy compared to single-feature models, with robust AUCs for 1-, 2-, and 3-year OS prediction across all cohorts. Risk stratification by the hybrid model revealed significant survival differences between low- and high-risk groups (all p < 0.05).

Conclusion: The proposed hybrid model, integrating MRI and radiomics features, enables accurate, non-invasive prediction of OS in CRCLM patients after targeted therapy, supporting individualized prognosis and clinical decision-making.

Trial registration: This study received the approval of the two Institutional Review Boards ( No.LLHBCH2025YN-082 for Hubei Cancer Hospital; No.UHCT-IEC-SOP-016-03-01 for Wuhan Union Hospital).

Colorectal cancer liver metastases

Targeted therapy

Prognosis

Radiomics

MRI

Prognostic model

Colorectal cancer (CRC) is one of the most prevalent malignancies globally. In 2023, it ranked third in both incidence and mortality across all cancer types, underscoring its substantial public health burden^[1]. The liver is the most common site of CRC metastasis, with approximately 20% of patients presenting with colorectal cancer liver metastases (CRCLM) at the time of initial diagnosis^[2].

For patients meeting surgical eligibility criteria—defined as resectable primary and metastatic lesions, adequate hepatic function, favorable general performance status, and absence of extrahepatic disease—hepatic resection remains the gold standard treatment. This approach confers the greatest potential for long-term survival in suitable candidates^[3]. In contrast, systemic chemotherapy serves as the cornerstone of management for unresectable CRCLM. The introduction of molecular targeted therapies has transformed the therapeutic landscape for these patients, facilitating tumor downstaging, increasing rates of conversion to resectability, and improving overall survival outcomes^[4–6].

The Response Evaluation Criteria in Solid Tumors (RECIST) remains the most widely adopted framework for assessing chemotherapy efficacy in solid malignancies^[7]. However, RECIST has inherent limitations: it relies solely on changes in tumor size, which may not accurately reflect the biological response to novel targeted agents. Additionally, it is susceptible to significant inter- and intra-observer variability, particularly in the assessment of irregularly shaped lesions.

Pathological evaluation using the tumor regression grade (TRG) system offers a more objective measure of treatment response and prognostic stratification^{[8, 9]}. Nevertheless, this approach is invasive, requires tissue sampling, and is prone to sampling bias—factors that restrict its utility for serial monitoring. Given these shortcomings, there is an urgent need for non-invasive, robust imaging biomarkers to guide prognostic assessment and treatment decision-making in CRCLM.

Magnetic resonance imaging (MRI) provides comprehensive tissue characterization beyond mere tumor size, including assessment of enhancement patterns and diffusion-weighted imaging (DWI)-derived apparent diffusion coefficient (ADC) metrics. Recent investigations have validated the prognostic value of delayed enhancement and ADC parameters in CRCLM^{[10, 11]}. Despite these advantages, these MRI-based approaches necessitate precise region-of-interest (ROI) delineation—a time-intensive process that limits its routine clinical utility. More importantly, they fail to capture the dynamic biological alterations and full spectrum of tumor microenvironmental heterogeneity that occur during treatment, creating a critical gap in personalized patient management.

Radiomics addresses these limitations by extracting high-dimensional quantitative features from standard imaging datasets, enabling objective characterization of tumor heterogeneity and microenvironmental properties_[12,13]. While most prior radiomics studies in CRCLM have focused on baseline intratumoral features, emerging evidence highlights the incremental prognostic value of peritumoral features and delta (temporal change) features^[14–16]. Specifically, delta radiomics— which quantifies longitudinal changes in imaging features over the course of treatment—has demonstrated improved performance in predicting treatment response and survival across various cancer types, including CRCLM. This dynamic assessment has the potential to inform clinical practice by detecting subtle biological changes (e.g., early signs of treatment resistance) before overt alterations in tumor size become apparent, thereby enabling timely adjustment of therapeutic strategies.

Building on these observations, we hypothesize that a delta radiomics model integrating both intratumoral and peritumoral regions will enhance the accuracy of overall survival (OS) prognostication in CRCLM patients receiving targeted therapy. The primary objective of this study is to develop and validate such a model. If successful, this tool could serve as a non-invasive adjunct to clinical practice, supporting individualized decision-making—such as stratifying patients into risk groups for personalized monitoring, guiding decisions on treatment continuation or switching, and refining assessments of resectability following neoadjuvant targeted therapy.

2.1 Patients

This retrospective study was conducted at two institutions: Hubei Cancer Hospital (Center 1) and Union Hospital, Wuhan (Center 2). Eligible patients were those with colorectal cancer (CRC) who received targeted therapy between January 2018 and December 2022 at Center 1, and between January 2020 and December 2022 at Center 2.

Inclusion criteria(Figure 1) were defined as: (1) pathologically confirmed colorectal adenocarcinoma; (2) complete clinical, pathological, and follow-up records available; (3) presence of liver metastases at initial diagnosis (confirmed by contrast-enhanced imaging or pathology); (4) availability of baseline liver MRI (performed within 1 month before initiating targeted therapy) and follow-up liver MRI (performed within 3 months after starting targeted therapy), with sequences adequate for radiomic analysis (including T1-weighted contrast-enhanced and T2-weighted sequences).

Exclusion criteria included: (1) incomplete or unevaluable MRI datasets (e.g., missing key sequences, insufficient image quality for lesion segmentation); (2) history of prior immunotherapy or local treatments targeting liver metastases (e.g., radiofrequency ablation, transarterial chemoembolization, or radiotherapy) before the initiation of targeted therapy; (3) presence of extrahepatic metastases (except for regional lymph nodes, which were not considered exclusionary); (4) target liver lesions <1 cm in maximum diameter (per RECIST 1.1 criteria, as smaller lesions are prone to measurement bias and unreliable feature extraction); (5) significant image artifacts (e.g., motion, breathing, or susceptibility artifacts) precluding accurate lesion segmentation or feature calculation.

For patients with multiple liver metastases, the three largest target lesions (per RECIST 1.1 guidelines) were selected for analysis to align with standard clinical response assessment practices. Treatment regimens are detailed in Supplementary Table 1.

2.2 MRI Imaging Protocol

Liver MRI was performed using standardized protocols on multiple 3.0T and 1.5T scanners across both centers. Sequences included T2-weighted imaging (T2WI), T1-weighted imaging (T1WI), and contrast-enhanced T1WI. Gadopentetate dimeglumine was used as the contrast agent. Detailed scanning parameters are provided in Supplementary Table 2.

2.3 Clinical Data Collection and Imaging Analysis

Clinical characteristics potentially associated with patient prognosis were retrospectively collected, including gender, age, and baseline carcinoembryonic antigen (CEA) and glycocalyx antigen 19-9 (CA19-9). Tumor markers were categorized into two groups based on serum levels: within normal range or above normal range (normal range: CEA ≤ 5 ng/ml, CA19-9 ≤ 40 ng/ml).

Multi-sequence MRI images were jointly evaluated by two physicians (RZ.X and W.X, with 5 and 3 years of abdominal imaging diagnosis experience, respectively). In case of discrepancies, a third physician (Physician 3, with 20 years of abdominal imaging experience) made the final decision. Up to three largest lesions per patient were evaluated for the following imaging features:(a) Number of liver metastases (≤ 3 or > 3); (b) Diameter (cm); (c) Maximum cross-sectional area (cm²); (d) Volume (cm³); (e) Growth pattern^[17]; (f) Metastasis homogeneity (homogeneous/heterogeneous); (g) Hepatic capsular retraction status; (h) Tumor related to adjacent vein (TRTAV); (i) Prominence of rim enhancement; (j) Clarity of boundary with surrounding liver parenchyma; (k) Regularity of lesion morphology.

Imaging Feature Assessment Principles(Figure2):

Diameter, maximum cross-sectional area, and volume were obtained by manual 3D ROI segmentation of the entire tumor using 3D Slicer software (Version 5.3.0, open-source software, https://www.slicer.org/) by two radiologists (RZ.X and W.X), with calculations performed by the software.
Growth patterns were classified into four types referring to Cai^[17]:Smooth Type (ST): Round or round-like tumor with a definite boundary and without protrusion at the edge of the lesion; Rough Type (RT): Tumor with multiple sharp-angled protrusions at the edge of the lesion, without a clear boundary with the surrounding liver parenchyma; Focal Extranodular Protuberant (FEP): Non-nodular tumor with one or more local protrusions at the edge of the lesion; Nodular Confluent (NC): Multiple nodules fused with each other, and each nodule had a clear outline.
Homogeneous: Signal intensity variation across the tumor (excluding necrotic cores) ≤10% of the mean intensity, as measured by 3D Slicer’s built-in intensity histogram tool. Visually, no distinct areas of hypo- or hyper-enhancement (relative to the tumor’s mean signal) were observed.

Heterogeneous: Signal intensity variation >10% of the mean intensity, with visually distinct regions of hypo-enhancement (e.g., necrosis) or hyper-enhancement (e.g., viable tumor).

Hepatic capsular retraction: Focal indentation of the liver capsule caused by the tumor, identified on PVP or T2-weighted images.
Positive TRTAV: Defined as (1) direct contact between the tumor and portal/hepatic vein trunks or their ≥2nd-order branches (distance between tumor margin and vessel wall <1 mm); or (2) visible intraluminal tumor thrombus or tumor vessels within the vein lumen, confirmed by lack of contrast enhancement in the affected segment.

Negative TRTAV: No contact with major veins, or distance between tumor and vessel ≥1 mm without intraluminal involvement.Prominent rim enhancement was defined as >75% of the tumor edge showing signal intensity significantly higher than surrounding liver parenchyma on axial images.

Clear boundary: ≥75% of the tumor circumference shows a distinct demarcation between the tumor and adjacent liver parenchyma, with no blurring or intermingling of signal intensities.
Regular morphology: Tumor contour approximates an ellipse, with no protrusions >5 mm in height or depressions >3 mm in depth relative to the best-fit ellipse .

2.4 Delineation of ROI and Extraction of Radiomic Features

The workflow of radiomic feature extraction is shown in Figure 3. ROIs were delineated on portal venous phase images of contrast-enhanced MRI at baseline and the first follow-up after targeted therapy. For patients with multiple lesions, up to three largest lesions (by diameter) were selected for ROI delineation. Two radiologists (RZ.X and W.X) manually segmented 3D ROIs of the entire tumor; peritumoral ROIsweregenerated by expanding the original ROI by 3 mm and manually correctingfor extrahepatic structures.

2.5 Image Preprocessing

All original MRI images underwent preprocessing to minimize central effects from different institutions and scanners. N4 bias field correction was performed to reduce image intensity inhomogeneity caused by magnetic field non-uniformity. To standardize voxel spacing and enhance texture resolution:

1.All ROIs were normalized using μ±3σ (μ = mean intensity within ROI; σ = standard deviation)

2.Gray-scale quantization was applied to reduce computation time and improve signal-to-noise ratio of texture results

3.Trilinear interpolation was used to isotropically resample images to 1×1×1 mm (x, y, z) voxel dimensions, preserving the proportion and orientation of 3D features

2.6 Feature Extraction

Features were extracted from intra-tumoral ROI (ROI_Intra), peri-tumoral ROI (ROI_Peri), and combined intra-peritumoral ROI (ROI_Combined) before and after targeted therapy using the Python Radiomics package. Delta radiomic features were calculated as: Δf_ti=f_ti-f_t0, where f_ti denotes the feature value at time point ti, and f_t0 denotes the baseline feature value.

2.7 Reliability Assessment

MRI data of 30 patients were randomly selected to calculate intra-observer and inter-observer correlation coefficients (ICC) for evaluating feature reliability. Inter-observer consistency was assessed by independent ROI segmentation of two radiologists at the same period. Intra-observer reproducibility was evaluated by repeated segmentation of one radiologist (RZ.X) two weeks later. Features with ICC > 0.75 were selected for further analysis (Supplementary Table 3).

2.8 Model Building

Before feature selection, the patient features were normalized using the Z-score formula: (x-μ)/σ, where x denotes the feature value, μ is the mean of all patient features, and σ is the corresponding standard deviation. The training cohort adopted a two-step feature selection approach.

First, Levene's test was performed to assess the homogeneity of variance for featuresand features with p-values > 0.05 were excluded. Second, the minimum redundancy maximum relevance (mRMR) method was used to select the top 5 features with high relevance and low redundancy.Finally, Six radiomics models (baseline and delta, intra/peri/combined) were constructed using logistic regression and 5-fold cross-validation. The best-performing model was selected. Univariate and multivariate Cox regression identified independent predictors of OS, which were used to build the final hybrid model.

2.8 Patient Follow-up

The overall survival (OS) of patients is defined as the time from initiation of targeted therapy to death or last follow-up.

2.9 Statistical Analysis

Statistical analyses were performed using Python (v3.9.6) and IBM SPSS Statistics 26.0. Continuous variables are presented as median (IQR); categorical variables as counts and percentages. Normality was tested using Kolmogorov-Smirnov test. Baseline characteristics were compared using Mann-Whitney U test/t-test or Chi-square test. Feature reliability was assessed using intraclass correlation coefficients (ICC > 0.75). Feature selection employed minimum redundancy maximum relevance (mRMR) algorithm. Model performance was evaluated using ROC/AUC with 95% confidence intervals, calibration curves, and decision curve analysis. Survival analysis used Kaplan-Meier method and Cox proportional hazards regression with proportional hazards assumption verification. Internal validation employed 5-fold cross-validation; external validation used independent cohort. SHAP analysis provided model interpretability. Statistical significance was set at p < 0.05.

3.1 Clinical Characteristics of Patients

After systematic application of the above criteria, 118 patients with colorectal cancer liver metastases (CRCLM) were included in the final cohort (83 from Center 1, 35 from Center 2).

Patients from Center 1 were randomly allocated to a training cohort and an internal test cohort at a 7:3 ratio using a computer-generated random number sequence, stratified by the number of liver metastases to ensure balanced distribution of key clinical characteristics. Patients from Center 2 were designated as an external validation cohort to evaluate the generalizability of the proposed model.

The demographic and clinical characteristics of patients in the training and test cohorts were comparable (Table 1). No significant differences in survival rates were observed between the training, test and validationsets (p=0.880, log-rank test; Figure 4). During follow-up, 32 patients (55%) in the training set and 14 (56%) in the test set died. The median OS was 30.5 months in the training set, 32.1 months in the test set and 26.0 months in the external validation set.The average OS was 32.4 months in the training set, 31.4 months in the test set and 22.3 months in the external validation set.

Table 1.Patient Characteristics
Variables	Training Set (n=58)	Testing Sets (n=25)	P-value
OS(month)	30.5(16.4-41.9)	32.1(18.5-43.0)	0.349
Age(year)	56.5(50.8-64.3)	55(51.5-65)	0.627
Diameter(cm)	2.89(1.9-4.9)	3.11(2.1-5.0)	0.747
Area(cm²)	5.345(2.6-14.8)	7.61(3.3-15.3)	0.595
Volume(cm³)	12.56(3.6-44.3)	16.5(5.2-40.5)	0.666
Sex			0.658
Male	40(69.0%)	16(64.0%)
Female	18(31.0%)	9(36.0%)
Number of liver metastases			0.311
≤3	23(39.7%)	7(28.0%)
＞3	35(60.3%)	18(72.0%)
Baseline CEA(ng/ml)			0.676
≤5	9(15.5%)	3(12.0%)
＞5	49(84.5%)	22(88.0%)
Baseline CA199(ng/ml)			0.567
≤40	15(25.9%)	8(32.0%)
＞40	43(74.1%)	17(68.0%)
ST			0.071
Y	42(72.4%)	13(52.0%)
N	16(27.6%)	12(48.0%)
RT			0.144
Y	27(46.6%)	16(64.0%)
N	31(53.4%)	9(36.0%)
FEP			0.710
Y	14(24.1%)	7(28.0%)
N	44(75.9%)	18(72.0%)
NC			0.825
Y	8(13.8%)	3(12.0%)
N	50(86.2%)	22(88.0%)
Hepatic capsular retraction status			0.507
Y	13(22.4%)	4(16.0%)
N	45(77.6%)	21(84.0%)
TRTAV			0.110
Y	36(62.1%)	20(80.0%)
N	22(37.9%)	5(20.0%)
Prominent rim enhancement			0.606
Y	36(62.1%)	17(68.0%)
N	22(37.9%)	8(32.0%)
Clear boundary			0.270
Y	45(77.6%)	22(88.0%)
N	13(22.4%)	3(12.0%)
Regular morphology			0.807
Y	45(77.6%)	20(80.0%)
N	13(22.4%)	5(20.0%)
Homogeneous lesions			0.567
Y	15(25.9%)	5(20.0%)
N	43(74.1%)	20(80.0%)

CEA：Carcinoembryonic Antigen；CA199：Glycocalyx Antigen 19-9；ST：Smooth Type；RT：Rough Type；FEP：Focal Extranodular Protuberant；NC：Nodular Confluent；TRTAV：Tumor related to adjacent vein

3.2Predictive Performance of Radiomics Models

A total of 1,688 features—including 14 shape, 324 first-order, and 1,350 texture features—were extracted from the intratumoral and peritumoral regions of interest (ROIs), with 3,376 features from the combined ROIs. Features with intraclass correlation coefficients (ICCs) below 0.75 were excluded. The top 5 features with high correlation and low redundancy were selected using t-tests and the minimum redundancy maximum relevance (mRMR) algorithm to construct the radiomics models. Details of feature selection and definitions are provided in Table 2.

Table 2.Composition of Feature in Different Radiomics Models
Rad-model	Composition of Feature
	wavelet-LHH_glrlm_LowGrayLevelRunEmphasis
	Log-sigma-4-0-mm-3D_glszm_LowGrayLevelZoneEmphasis
Model_intra	wavelet-HHH_glszm_HighGrayLevelZoneEmphasis
	Original_gldm_LargeDependencelLowGrayLevelEmphasis
	Wavelet-LLH_firstorder_90Percentile

	squareroot_firstorder_Kurtosis
	log-sigma-4-0-mm-3D_firstorder_RobustMeanAbsoluteDeviation
Model_peri	wavelet-HHL_glszm_SmallAreaLowGrayLevelEmphasis
	wavelet-HHL_firstorder_Skewness
	wavelet-LHH_glszm_HighGrayLevelZoneEmphasis

	I_wavelet-LHH_glrlm_LowGrayLevelRunEmphasis
	P_wavelet-HLL_firstorder_Minimum
Model_combined	I_wavelet-HHH_glszm_HighGrayLevelZoneEmphasis
	I_log-sigma-4-0-mm-3D_glszm_SmallAreaLowGrayLevelEmphasis
	P_wavelet-HHL_firstorder_Skewness

	ΔI_wavelet-HHH_glszm_LowGrayLevelZoneEmphasis
	ΔI_log-sigma-4-0-mm-3D_gldm_SmallDependenceHighGrayLevelEmphasis
ΔModel_intra	ΔI_original_glszm_ZoneVariance
	ΔI_wavelet-LLL_firstorder_Skewness
	ΔI_wavelet-HHH_firstorder_Skewness

	ΔP_wavelet-LHH_glrlm_HighGrayLevelRunEmphasis
	ΔP_wavelet-HHL_glszm_SizeZoneNonUniformityNormalized
ΔModel_peri	ΔP_log-sigma-2-0-mm-3D_firstorder_Range
	ΔP_wavelet-HLH_gldm_HighGrayLevelEmphasis
	ΔP_wavelet-LHL_glszm_GrayLevelVariance

	ΔI_wavelet-HHH_glszm_LowGrayLevelZoneEmphasis
	ΔI_log-sigma-4-0-mm-3D_gldm_SmallDependenceHighGrayLevelEmphasis
ΔModel_combined	ΔI_original_glszm_ZoneVariance
	ΔP_wavelet-LHH_glcm_SumAverage
	ΔI_wavelet-LLL_firstorder_Skewness

Figures 5 and 6 illustrate the ROC curves for baseline and delta radiomics models. In the training set, the AUCs for baseline intratumoral, peritumoral, and combined models were 0.845, 0.823, and 0.874, respectively; for delta models, 0.852, 0.854, and 0.857. In the test set, the AUCs for baseline models were 0.785, 0.740, and 0.828, and for delta models, 0.814, 0.761, and 0.855. Delta radiomics models outperformed their baseline counterparts, with the delta combined model showing the best predictive performance. Table 3 summarizes the accuracy, sensitivity, and specificity of all models. The SHAP plot (Figure 7) demonstrates the contribution of each feature in the delta combined model.

Table 3.Predictive Performance of Radiomics Models
	AUC	Accuracy	Sensitivity	Specificity
Training Sets (n=58)
Rad-model_intra	0.845	0.738	0.698	0.743
Rad-model_peri	0.823	0.756	0.649	0.732
Rad-model_combined	0.874	0.816	0.828	0.837
ΔRad-model_intra	0.852	0.774	0.731	0.773
ΔRad-model_peri	0.854	0.786	0.721	0.773
ΔRad-model_combined	0.857	0.777	0.718	0.772
Testing Sets (n=25)
Rad-model_intra	0.785	0.700	0.700	0.718
Rad-model_peri	0.740	0.699	0.680	0.669
Rad-model_combined	0.828	0.769	0.790	0.766
ΔRad-model_intra	0.814	0.726	0.697	0.741
ΔRad-model_peri	0.761	0.734	0.690	0.720
ΔRad-model_combined	0.855	0.760	0.760	0.783

3.3 Construction and Performance Testing of Hybrid Models

The Rad-score for each patient was calculated using the delta combined radiomics model. Univariate and multivariate Cox regression analyses identified FEP, TRTAV, and Rad-score as independent predictors of OS（Table 4）. The hybrid model, constructed from these variables(FEP, TRTAV and Rad-score), achieved AUCs of 0.925 (training), 0.782 (test), and 0.750 (external validation) for OS prediction (Figure 8).

Table 4. Cox regression analysis
Variables	UnivariateCox regression analyses		MultivariateCox regression analyses
Variables	P-value	HR(90%CI)	P-value	HR(95%CI)
Sex	0.175	0.601(0.324-1.115)
Age	0.297	1.021(0.988-1.054)
Number of liver metastases	0.305	0.681(0.368-1.260)
Baseline CEA(ng/ml)	0.468	0.676(0.278-1.641)
Baseline CA199(ng/ml)	0.429	0.723(0.368-1.420)
Diameter(cm)	0.973	0.997(0.874-1.138)
Area(cm²)	0.499	0.990(0.967-1.014)
Volume(cm³)	0.425	0.998(0.994-1.002)
ST	0.403	0.717(0.373-1.379)
RT	0.458	1.303(0.725-2.343)
FEP	0.054	2.115(1.115-4.012)	0.042	2.247(1.031-4.895)
NC	0.546	1.345(0.600-3.017)
Hepatic capsular retraction status	0.935	0.966(0.477-1.955)
TRTAV	0.045	2.133(1.145-3.971)	0.041	2.239(1.034-4.846)
Prominent rim enhancement	0.956	0.980(0.542-1.774)
Clear boundary	0.614	0.813(0.414-1.596)
Regular morphology	0.718	0.861(0.437-1.699)
Homogeneous lesions	0.154	0.524(0.249-1.105)
Rad-score	＜0.001	0.063(0.021-0.191)	＜0.001	0.05(0.12-0.214)

CEA：Carcinoembryonic Antigen；CA199：Glycocalyx Antigen 19-9；ST：Smooth Type；RT：Rough Type；FEP：Focal Extranodular Protuberant；NC：Nodular Confluent；TRTAV：Tumor related to adjacent vein

3.4 Visualization and Application of Hybrid Models

A nomogram was developed to visualize the hybrid model, providing individualized OS probability estimates (Figure 9). The nomogram's AUCs for 1-, 2-, and 3-year OS were 0.632, 0.738, and 0.873 in the training set; 0.561, 0.754, and 0.847 in the test set; and 0.560, 0.665, and 0.810 in the external validation set (Figure 10).Calibration curves confirmedgood agreement between the 3-year OS predicted by the hybrid model and the actual OS in all groups(Figure 11).

3.5 Comparison with the RECIST 1.1 criteria

According to RECIST 1.1, patients were classified as progressive disease (PD), partial response (PR), stable disease (SD), or complete response (CR). Kaplan-Meier analysis showed no significant OS difference between response and non-response groups by RECIST 1.1(training set : p=0.881; test set : p=0.116; external validation set : p=0.109).

After calculating individual total scores via the nomogram, the optimal cutoff was determined using the Youden index from the ROC curve of the training cohort, with this threshold applied consistently to all cohorts.Patients were divided into a low-risk group (total score ≤157) and a high-risk group (total score >157) based on the total score calculated by the nomogram. Kaplan-Meier survival analysis revealed significantly better OS in the low-risk group across all cohorts(training set : p＜0.00r01; test set : p=0.038; external validation set : p=0.037)(Figures 12–14).

A 61-year-old male with non-FEP lesions and TRTAV(-) had a total score of 155 points (low-risk group). According to the RECIST 1.1 criteria, the evaluation was progressive disease (PD, non-response group). The follow-up duration was 5.7 years, and the patient was alive.

A 55-year-old male with FEP lesions and TRTAV(+) had a total score of 227 points (high-risk group). According to the RECIST 1.1 criteria, the evaluation was partial response (PR, response group). The follow-up duration was 2.1 years, and the patient died.

Targeted therapies have revolutionized the management of colorectal cancer liver metastases (CRCLM), offering improved survival outcomes compared to traditional cytotoxic chemotherapy regimens^[18]. However, the current gold standard for treatment response assessment—the Response Evaluation Criteria in Solid Tumors (RECIST) 1.1—exhibits significant limitations in evaluating the efficacy of molecular targeted agents, which often induce cytostatic rather than cytotoxic effects^[7]. Our study addresses this critical gap by developing a hybrid prognostic model that integrates delta radiomics with conventional MRI features to predict overall survival (OS) in CRCLM patients undergoing targeted therapy. The model's superior performance compared to RECIST 1.1 criteria (AUC 0.925 vs. non-significant stratification) underscores the potential of radiomics-based approaches to transform clinical decision-making in this patient population.

The clinical implications of our findings extend beyond academic interest to direct patient care optimization. The hybrid model's ability to stratify patients into distinct risk groups (low-risk: total score ≤ 157; high-risk: total score > 157) with significantly different survival outcomes (p < 0.05) provides clinicians with a practical tool for individualized treatment planning. This stratification enables several critical clinical decisions: (1) identification of high-risk patients who may benefit from more aggressive treatment regimens or early consideration of alternative therapies; (2) optimization of follow-up schedules for low-risk patients, potentially reducing unnecessary imaging and healthcare costs; (3) informed discussions with patients and families regarding prognosis and treatment goals. The nomogram visualization further enhances clinical utility by providing intuitive, point-based risk assessment that can be readily integrated into routine clinical workflows.

Our findings align with and extend previous radiomics research in CRCLM while addressing several methodological gaps. Similar to the work of Ammirabile et al.^[19], who demonstrated MRI-based radiomics prediction of pathologic response in colorectal liver metastases, our study confirms the prognostic value of radiomics features in this patient population. However, our approach differs significantly in several key aspects. First, we incorporated delta radiomics—a temporal analysis approach that captures treatment-induced changes—whereas previous studies primarily focused on baseline features^{[20, 21]}. This temporal dimension proved crucial, as our delta combined model (AUC 0.855) outperformed baseline models (AUC 0.828), consistent with findings in other cancer types^{[22, 23]}.The prognostic significance of tumor margin characteristics in our study corroborates established literature while providing novel insights. Cai et al.^[17] previously demonstrated that non-smooth margins (FEP and NC types) are associated with poorer disease-free survival in CRCLM. Our findings extend this observation to overall survival prediction in the targeted therapy setting, with FEP lesions showing a 2.247-fold increased risk of death (95% CI: 1.031–4.895, p = 0.042). This consistency across different treatment modalities suggests that tumor morphology reflects fundamental biological properties that transcend specific therapeutic approaches.

The delta radiomics approach represents a paradigm shift from static to dynamic assessment of tumor response. Our identification of ΔI_wavelet-LLL_firstorder_Skewness as a key prognostic feature(for the average SHAP plot showed thatthe delta feature "ΔI_wavelet-LLL_firstorder_Skewness" ranked among the top two in terms of contribution in all 4 folds of the training model) provides novel insights into the biological mechanisms underlying treatment response. Skewness changes in medical imaging reflect alterations in tissue composition and vascular architecture^{[24, 25]}. In the context of targeted therapy, a smaller decrease in skewness post-treatment suggests limited tumor necrosis—a marker of inadequate therapeutic response. This finding aligns with established knowledge that effective targeted therapy should induce significant tumor cell death and subsequent necrosis^[26].The clinical relevance of delta radiomics extends beyond prognostic prediction to real-time treatment monitoring. Unlike RECIST 1.1, which requires substantial size changes to detect response, delta radiomics can identify subtle biological changes before morphological alterations become apparent^[27]. This early detection capability is particularly valuable in targeted therapy, where treatment resistance often develops gradually and early intervention can significantly impact patient outcomes. Our model's ability to predict 1-, 2-, and 3-year OS with robust accuracy (AUCs: 0.632–0.873) provides clinicians with a tool for both immediate risk assessment and long-term prognosis planning.

In most radiomics studies^[28–30], feature selection and model construction typically involve a two-step process: first, features are selected using the minimum redundancy maximum relevance (mRMR) method, and then further selection and modeling are performed using the Least Absolute Shrinkage and Selection Operator (LASSO). LASSO primarily incorporates an L1 regularization term into the loss function, causing the coefficients of some features to shrink to zero, thereby achieving feature selection. Additionally, LASSO can function as a classifier; essentially, it is a logistic regression model that can use the selected features to build a logistic regression model and output predicted probabilities for each sample.Multiple rounds of feature screening and dimensionality reduction for a large number of features can effectively avoid overfitting in the final model. However, in this study, severe overfitting still occurred even after using LASSO or other machine learning methods (Supplementary Table 5). One possible reason is the small sample size of the training set in this study, which made it difficult for the model to accurately represent the true data distribution, thus easily overfitting to details and noise in the training data^{[30, 31]}. Noise in the training data and biases in data distribution may also cause overfitting, as biased data may not represent the true data distribution, leading the model to over-adapt to these biases during training.Therefore, this study only used mRMR for feature screening. To avoid overfitting, only the top 5 features with high relevance and low redundancy were selected, rather than the top 20 features as in previous studies^{[28, 32]}.

However, this study still has some limitations: (1) The multi-center design, while enhancing generalizability, introduced inherent variability in scanner types, acquisition parameters, and imaging protocols across institutions. Although we attempted to standardize preprocessing using N4 bias field correction and isotropic resampling^[33], subtle differences in image quality and contrast enhancement timing may have influenced feature extraction. Future studies should implement more robust harmonization techniques, such as ComBat correction or deep learning-based domain adaptation, to further reduce inter-site variability^{[34, 35]}.(2) The manual ROI delineation process, while ensuring clinical relevance, introduces potential inter-observer variability. Our ICC analysis demonstrated good to excellent consistency (ICC > 0.75) for selected features, but the process remains time-intensive and subject to human error. Emerging automated segmentation techniques, particularly deep learning-based approaches, may address this limitation while maintaining clinical accuracy^[36].(3) Selection bias represents a primary concern, as patients included in our study may not represent the broader CRCLM population. Our exclusion criteria, while necessary for methodological rigor, may have inadvertently selected for patients with more favorable baseline characteristics or better treatment adherence. The relatively small sample size, particularly in the external validation cohort (n = 35), further limits the statistical power and generalizability of our findings. Power analysis suggests that our study was underpowered to detect small but clinically meaningful differences in survival outcomes.The limited pathological correlation represents another significant limitation of our retrospective design. With only 2 pathological specimens available for analysis, we were unable to directly validate the biological mechanisms underlying our radiomics findings. This limitation is particularly relevant for delta radiomics interpretation, where pathological correlation could provide crucial insights into the relationship between imaging changes and underlying tumor biology.

In conclusion, our study demonstrates that a hybrid model integrating delta radiomics with conventional MRI features can accurately predict overall survival in CRCLM patients undergoing targeted therapy. The model's superior performance compared to RECIST 1.1 criteria, combined with its practical nomogram visualization, positions it as a valuable tool for individualized patient management. However, the retrospective design, limited cohort size, and imaging variability challenges underscore the need for prospective validation and methodological refinement. Future research should focus on addressing these limitations while advancing the clinical translation of radiomics-based prognostic tools in CRCLM management.

List of abbreviations
ADC	Apparent diffusion coefficient
AUC	Area under the curve
CA199	Glycocalyx antigen 19-9
CEA	Carcinoembryonic antigen
CI	Confidence interval
CR	Complete response
CRC	Colorectal cancer
CRCLM	Colorectal cancer liver metastases
CT	Computer tomography
FEP	Focal extranodular protuberant
HGPs	Histopathological growth patterns
ICC	Intra/Inter-observer correlation cofficients
MRI	Magnetic resonance imaging
mRMR	minimum Redundancy maximum relevance
MVI	Microvascular invasion
NC	Nodular confluent
OS	Overall survival
PD	Progressive disease
PFS	Progression-free survial
PR	Partial response
RECIST	Response evaluation criteria solid tumors
ROC	Receiver operating characteristic curve
ROI	Region of interesting
RT	Rough type
SD	Stable disease
SHAP	SHapley Additive exPlanation
ST	Smooth type
TE	Echo time
TR	Repetition time
TRG	Tumor regression grade
TTPVI	Two-trait predictor of venous invasion
TRTAV	Tumor related to adjacent vein
VEGF	Vascular endothelial growth factor

Ethics approval and consent to participate

Written informed consent was obtained from all patients recruited in this study. All methods were carried out in accordance with Declaration of Helsinki and Good Clinical Practice (GCP) guidelines. Institutional Review Board of Hubei Cancer Hospital and Wuhan Union Hospital have approved the study protocol. This study received the approval of the two Institutional Review Boards ( No.LLHBCH2025YN-082 for Hubei Cancer Hospital; No.UHCT-IEC-SOP-016-03-01 for Wuhan Union Hospital).

Consent for publication

Written form of consent for publication have been obtained from all of the patients whom involved in this study.

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author by e-mail on reasonable request.

Competing interests

The authors declare that they have no conflict of interest.

Funding

Y.L. was supported by Talent Project of Hubei Cancer Hospital (NO:2025HBCHLHRC001), Chutian Talents of Hubei (NO: CTYC001), Hubei Provincial Key Technology Foundation of China (grant no. 2021ACA013). Y.W. was supported by the Talent Project of Hubei Cancer Hospital (grant no. 2025HBCHHHRC006).

Authors’ contributions

Renzhe Xiao: Conceptualization, Methodology, Formal analysis, Investigation, Writing-Original Draft

Shuangquan Ai: Software, Validation, Visualization

Yi Li: Data Curation

Wei Xiao: Data Curation

Zixuan Liu: Data Curation

Xiaofang Guo: Writing-Review & Editing, Supervision

Yulin Liu: Project administration, Funding acquisition

Acknowledgements

Not applicable.

Siegel R L, Wagle N S, Cercek A, et al. Colorectal cancer statistics, 2023[J]. CA: a cancer journal for clinicians, 2023, 73(3): 233–254.
Chinese College of Surgeons; Section of Gastrointestinal Surgery, et al. China guideline for diagnosis and comprehensive treatment of colorectal liver metastases (version 2023). Zhonghua Wei Chang Wai Ke Za Zhi. 2023 Jan 25;26(1):1-15.
Wang F, Chen G, Zhang Z, et al. The Chinese Society of Clinical Oncology (CSCO): Clinical guidelines for the diagnosis and treatment of colorectal cancer, 2024 update. Cancer Commun (Lond). 2025 Mar;45(3):332-379.
BORASCHI P, DONATI F, CERVELLI R, et al. Colorectal liver metastases:ADC as an imaging biomarker of tumor behavior and therapeutic response[J/OL]. Eur J Radiol, 2021, 137:109609.
DONATI F, BORASCHI P, PACCIARDI F, et al. 3T diffusion-weighted MRI in the response assessment of colorectal liver metastases after chemotherapy:correlation between ADC value and histological tumour regression grading[J]. Eur J Radiol, 2017, 91:57-65.
HOSSEINI-NIK H, FISCHER S E, MOULTON C A, et al.Diffusion-weighted and hepatobiliary phase gadoxetic acid-enhanced quantitative MR imaging for identification of complete pathologic response in colorectal liver metastases after preoperative chemotherapy[J]. Abdom Radiol, 2016, 41(2):231-238.
Eisenhauer EA, Therasse P, Bogaerts J, et al. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1)[J]. Eur J cancer (Oxford), 2009, 45(2): 228-247.
Cai Y, Lu X, Zhu X, et al. Histological tumor response assessment in colorectal liver metastases after neoadjuvant chemotherapy: impact of the variation in tumor regression grading and peritumoral lymphocytic infiltration. J Cancer. 2019 Oct 6;10(23):5852-5861.
Serayssol C, Maulat C, Breibach F, et al. Predictive factors of histological response of colorectal liver metastases after neoadjuvant chemotherapy. World J Gastrointest Oncol. 2019 Apr 15;11(4):295-309.
Cheung HMC, Karanicolas PJ, Coburn N, et al. Delayed tumour enhancement on gadoxetate-enhanced MRI is associated with overall survival in patients with colorectal liver metastases. Eur Radiol. 2019 Feb;29(2):1032-1038.
Zhu HB, Xu D, Zhang XY, et al. Prediction of Therapeutic Effect to Treatment in Patients with Colorectal Liver Metastases Using Functional Magnetic Resonance Imaging and RECIST Criteria: A Pilot Study in Comparison between Bevacizumab-Containing Chemotherapy and Standard Chemotherapy. Ann Surg Oncol. 2022 Jun;29(6):3938-3949.
Lo Gullo R, Marcus E, Huayanay J, et al. Artificial intelligence-enhancedbreast MRI: Applications inbreast cancer primary treatment responseassessment and prediction. Invest Radiol 2023. https://doi.org/10.1097/RLI.000000000000 1010.
Bian T, Wu Z, Lin Q, et al. Evaluating tumor-infiltrating lymphocytes inbreast cancer using preoperative MRI-based radiomics. J Magn ResonImaging 2022;55(3):772-784.
Braman NM, Etesami M, Prasanna P, et al. Intratumoral and peritumoralradiomics for the pretreatment prediction of pathological completeresponse to neoadjuvant chemotherapy based on breast DCE-MRI.Breast Cancer Res 2017;19:57.
Khorrami M, Prasanna P, Gupta A, et al. Changes in CT Radiomic Features Associated with Lymphocyte Distribution Predict Overall Survival and Response to Immunotherapy in Non-Small Cell Lung Cancer. Cancer Immunol Res 2020;8(1):108–119.
Xia TY, Zhou ZH, Meng XP, et al. Predicting Microvascular Invasion in Hepatocellular Carcinoma Using CT-based Radiomics Model. Radiology. 2023 May;307(4):e222729. doi: 10.1148/radiol.222729.
Cai Q, Mao Y, Dai S, et al. The growth pattern of liver metastases on MRI predicts early recurrence inpatients with colorectal cancer: a multicenter study. Eur Radiol. 2022 Nov;32(11):7872-7882.
Underwood PW, Ruff SM, Pawlik TM. Update on Targeted Therapy and Immunotherapy for Metastatic Colorectal Cancer. Cells. 2024 Jan 28;13(3):245. doi: 10.3390/cells13030245.
Ammirabile A, Levi R, Boldrini L, et al. MRI-based radiomics predicts the pathologic response of colorectal liver metastases to systemic therapy: A multicenter study. Eur J Radiol. 2025;192:112325.
Zhang B, He X, Ouyang F, et al. Radiomic machine-learning classifiers for prognostic biomarkers of advanced nasopharyngeal carcinoma. Cancer Lett. 2017;403:21-27.
Fave X, Zhang L, Yang J, et al. Delta-radiomics features for the prediction of patient outcomes in non-small cell lung cancer. Sci Rep. 2017;7(1):588.
Chang Y, Lafata K, Sun W, et al. An investigation of machine learning methods in delta-radiomics feature analysis. PLoS One. 2019;14(12):e0226348.
Wu J, Li B, Sun X, et al. Heterogeneous enhancement patterns of tumor-adjacent parenchyma at MR imaging are associated with dysregulated signaling pathways and poor survival in breast cancer. Radiology. 2017;285(2):401-413.
Enkhbaatar NE, Inoue S, Yamamuro H, et al. MR Imaging with Apparent Diffusion Coefficient Histogram Analysis: Evaluation of Locally Advanced Rectal Cancer after Chemotherapy and Radiation Therapy. Radiology. 2018;288(1):129-137.
Baek HJ, Kim HS, Kim N, et al. Percent change of perfusion skewness and kurtosis: a potential imaging biomarker for early treatment response in patients with newly diagnosed glioblastomas. Radiology. 2012;264(3):834-843.
Underwood PW, Ruff SM, Pawlik TM. Update on Targeted Therapy and Immunotherapy for Metastatic Colorectal Cancer. Cells. 2024;13(3):245.
Lo Gullo R, Marcus E, Huayanay J, et al. Artificial intelligence-enhanced breast MRI: Applications in breast cancer primary treatment response assessment and prediction. Invest Radiol. 2023;58(8):505-515.
Xu Y, Ye F, Li L, Yang Y, Ouyang J, Zhou Y, Wang S, Xie L, Zhou J, Zhao H, Zhao X. MRI-Based Radiomics Nomogram for Preoperatively Differentiating Intrahepatic Mass-Forming Cholangiocarcinoma From Resectable Colorectal Liver Metastases. Acad Radiol. 2023 Sep;30(9):2010-2020.
Wu Z, Lin Q, Wang H, et al. Intratumoral and Peritumoral Radiomics Based on Preoperative MRI for Evaluation of Programmed Cell Death Ligand-1 Expression in Breast Cancer. J Magn Reson Imaging. 2024 Aug;60(2):588-599.
Bejani M M , Ghatee M .A systematic review on overfitting control in shallow and deep neural networks[J].Artificial Intelligence Review, 2021:1-48.DOI: 10.1007/s10462-021-09975-1.
Ying X .An Overview of Overfitting and its Solutions[J].Journal of Physics: ConferenceSeries, 2019, 1168:022022-. DOI:10.1088/1742-6596/1168/2/ 022022.
Xia TY, Zhou ZH, Meng XP, et al. Predicting Microvascular Invasion in Hepatocellular Carcinoma Using CT-based Radiomics Model. Radiology. 2023 May;307(4):e222729.
Orlhac F, Frouin F, Nioche C, et al. Validation of a method to compensate multicenter effects affecting CT radiomics. Radiology. 2019;291(1):53-59.
Fortin JP, Cullen N, Sheline YI, et al. Harmonization of cortical thickness measurements across scanners and sites. Neuroimage. 2017;167:104-120.
Zuo L, Zhang G, Li Z, et al. Deep learning-based radiomics for predicting pathological complete response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Eur Radiol. 2021;31(12):9319-9328.
Nicoli AP, Bach M, Wasserthal J, Indrakanti AK, Segeroth M, Yang S, Cyriac J, Boll D, Wilder-Smith AJ. Liver Segment and Lesion Segmentation on CT and MRI: An Open-Source Contribution to TotalSegmentator. J Imaging Inform Med. 2025 Oct 24.

No competing interests reported.

Appendix.docx

Intratumoral and peritumoral delta radiomics of MRI predicts overall survival to targeted therapy in colorectal cancer with liver metastases

Status:

Version 1

Abstract

Figures

1 Background

2 Methods

2.1 Patients

2.2 MRI Imaging Protocol

2.3 Clinical Data Collection and Imaging Analysis

2.4 Delineation of ROI and Extraction of Radiomic Features

2.5 Image Preprocessing

2.6 Feature Extraction

2.7 Reliability Assessment

2.8 Model Building

2.8 Patient Follow-up

2.9 Statistical Analysis

3 Result

3.1 Clinical Characteristics of Patients

3.2Predictive Performance of Radiomics Models

3.3 Construction and Performance Testing of Hybrid Models

3.4 Visualization and Application of Hybrid Models

3.5 Comparison with the RECIST 1.1 criteria

4 Discussion

5 Conclusion

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1