Preoperative magnetic resonance imaging-based radiomics nomogram model for predicting postoperative anastomotic leakage in colorectal cancer.

INTRODUCTION
Anastomotic leakage (AL) remains one of the most feared and devastating complications following colorectal cancer surgery. AL adversely affects surgical outcomes, prolongs hospital stays, increases healthcare costs, raises postoperative tumor recurrence rates, and reduces both overall and cancer-specific survival [1, 2]. Despite significant advancements in perioperative care, surgical techniques, and instrumentation, the incidence of AL has remained steady at 3% to 15% over the past several decades [3, 4]. For some patients, AL appears almost unavoidable, underscoring the importance of early identification and prevention. Numerous studies have reported various risk and predictive factors for AL, including tumor size, tumor height, sex, nutritional status, body weight, smoking, comorbid diabetes, and C-reactive protein levels [5–7]. Although many of these factors are unmodifiable, strategies such as prophylactic stoma creation, reinforcement of the anastomosis with additional sutures, and modifications in anastomotic technique may reduce or prevent AL [8–10].
Diagnosing AL is also challenging, as clinical manifestations range from asymptomatic presentations to life-threatening septic shock. Delayed diagnosis can result in delayed treatment and worse outcomes [11, 12]. Therefore, a reliable preoperative method for predicting AL is critically needed to facilitate timely intervention and preventive strategies.
Radiomics, an emerging technique, quantitatively extracts imaging features to evaluate tumor heterogeneity and biological behavior. It holds significant promise in personalized medicine for rectal cancer by improving diagnosis, guiding treatment, and predicting prognosis [13, 14]. Recent studies have increasingly applied radiomics for tumor diagnosis, therapeutic planning, and prognostic assessment, demonstrating its diagnostic and predictive value [15, 16]. Radiomics provides a comprehensive evaluation of tumor heterogeneity and biological behavior. However, only a few studies to date have explored its application for predicting AL in colorectal cancer. Thus, the present study aims to develop and validate a radiomics-based predictive model for the preoperative assessment of AL.

METHODS

Ethics statement
This study was approved by the Research Ethics Committee of Renmin Hospital of Wuhan University (No. WDRY2018-K055). Informed consent was waived due to the use of deidentified data and the retrospective nature of the study.

Patients
Patients who underwent surgery for colorectal cancer at the Gastrointestinal Surgery Center of Renmin Hospital of Wuhan University between January 2018 and December 2024 were screened for inclusion. All enrolled patients underwent colorectal cancer surgery and had complete clinical data, including magnetic resonance imaging (MRI) within 1 week prior to surgery. Inclusion criteria were as follows: (1) colorectal adenocarcinoma confirmed by pathology; (2) complete clinical data; and (3) preoperative MRI, including T1-weighted imaging (T1WI) and T2WI, performed within 14 days prior to surgery. Exclusion criteria were as follows: (1) nonprimary adenocarcinoma pathology; (2) incomplete clinical data or suboptimal image quality for analysis; and (3) patients who underwent Hartmann procedure, Miles procedure, or other procedures without intestinal anastomosis. Eligible patients were randomly divided into training and validation sets at a 7:3 ratio (Fig. 1). Two radiologists, with 5 and 7 years of experience respectively, independently analyzed MRI features without knowledge of postoperative outcomes, and their assessments were consistent.
A total of 146 patients were included in our study. Clinical characteristics, including age, sex, body mass index, and protective ostomy were obtained from hospital records. Postoperative AL was defined according to the International Study Group of Rectal Cancer (ISREC) criteria [12], including grade B and C leaks confirmed through clinical, radiological, or surgical evidence. All patients were followed for a minimum of 30 days postoperatively, during which all observed AL events were documented. Baseline patient data are detailed in Table 1.

Development of the clinical model
Clinical data collected included age, sex, body mass index, diabetes, neoadjuvant therapy, albumin levels, and distance from the tumor to the anal verge. Odds ratios (ORs) and 95% confidence intervals (CIs) were calculated for each factor. Based on univariate analysis, age, tumor distance from the anal verge, and neoadjuvant therapy showed significant differences between with and without AL groups (P<0.2). These 3 factors were subsequently incorporated into a multivariate logistic regression model, which constituted the clinical model.

Image segmentation, preprocessing, and radiomic feature extraction
Preprocessing steps were performed to standardize image quality and intensity. First, N4ITK bias field correction was applied to adjust for intensity inhomogeneity caused by scanner magnetic field variations during image acquisition [17, 18]. All images were then resampled to an isotropic voxel size of 1×1×1 mm3 to ensure spatial uniformity across datasets [19]. Finally, Z-score normalization was applied to the extracted features to minimize scale differences and improve reproducibility.
Without knowledge of clinical outcomes, 2 radiologists manually delineated the tumor boundaries on T1WI and fat-suppressed T2WI (fs-T2WI) using 3D Slicer ver. 5.6.2 (Slicer Community; https://www.slicer.org) (Fig. 2). Adjacent normal tissue and blood vessels were excluded. From the region of interest on the T1WI and fs-T2WI sequences, a total of 1,702 3-dimensional MR radiomic features were extracted.
To assess reproducibility, interobserver consistency was tested by randomly selecting 10 MRI images (5 with AL and 5 without AL). Two radiologists independently performed region of interest segmentation on both T1WI and fs-T2WI images. The intraclass correlation coefficient (ICC) was used to evaluate consistency, with ICC >0.75 indicating good agreement. Radiologist 1 then extracted radiomic features from the remaining images.

Feature selection and development of the radiomics model
Only features with ICC >0.75 were retained for further analysis. The least absolute shrinkage and selection operator (LASSO) algorithm was applied to the training set to identify optimal features. LASSO regression, a widely used shrinkage technique, reduces overfitting and improves predictive accuracy [20]. Radiomic features selected by LASSO were used to construct the radiomics model. A radiomic score (Rad score) was then calculated for each patient using a linear combination of selected feature values (A [wavelet.LLHglcmDifferenceVariance], measures the variance of gray-level differences in the gray-level co-occurrence matrix [GLCM], indicating texture variation and image detail; B [wavelet.LHLglcmJointEnergy], reflects the joint energy in the GLCM, reflecting the uniformity or simplicity of the texture; C [wavelet.HHHglcmImc1], based on informational measure of correlation 1 (IMC1) of the GLCM, it indicates the texture dependency, with higher values showing more regular textures; D [wavelet.HHHgldmSmallDependenceHighGrayLevelEmphasis], emphasizes small regions with high gray levels in the gray-level dependence matrix [GLDM], reflecting areas of high contrast; E [wavelet.HHHgldmSmallDependenceLowGrayLevelEmphasis], emphasizes small regions with low gray levels in the GLDM, highlighting low-contrast areas; F [originalngtdmStrength], measures the strength of the texture using the neighborhood gray-tone difference matrix [NGTDM], with higher values indicating stronger and more intricate textures; G [wavelet.LHLglcmImc1], similar to IMC1 of the GLCM, calculated from the LHL subband, reflecting texture complexity; H [wavelet.LHLglcmImc2], based on IMC2 of the GLCM, it reflects finer texture details and contrast in the image; I [wavelet.LHLgldmSmallDependenceEmphasis], measures the importance of small dependent regions in the image using the GLDM, indicating texture relevance) weighted by their corresponding nonzero coefficients. The Rad score was calculated using the following formula:

Development of the radiomics nomogram model
A radiomics nomogram model incorporating the Rad score and clinical factors selected by the clinical model (including age, tumor distance from the anal verge, and neoadjuvant therapy) was constructed using multivariable logistic regression. Supplementary Table 1 provides the regression coefficients and ORs corresponding to each included variable. Subsequently, a nomogram score was generated for each patient based on the model.

Model evaluation
The area under the curve (AUC) of the receiver operating characteristic (ROC) curve was calculated for both training and validation sets to assess the discriminatory performance of the clinical model, the radiomics model, and the radiomics nomogram model. In addition, decision curve analysis was performed to evaluate the clinical utility of the 3 models by quantifying their net benefit across different threshold probabilities in the validation set.

Statistical analysis
Statistical analysis was performed using SPSS ver. 17.0 (SPSS Inc) and R ver. 4.4.1 (R Foundation for Statistical Computing). A 2-sided P-value of <0.05 was considered statistically significant. The chi-square test and Fisher exact test were used for categorical variables, and the independent samples t-test was used for continuous variables. The following R packages were applied: “pROC” for ROC curve analysis, “rms” for nomograms and calibration curves, “glmnet” for LASSO logistic regression, and “rmda” for decision curve analysis.

RESULTS

Development of the clinical model
A total of 146 patients were included in this study, among whom 11 (7.5%) developed postoperative AL. Table 1 summarizes the clinical characteristics of patients, stratified into with and without AL groups. ORs and 95% CIs were calculated for each independent factor. Univariate analysis indicated that age, tumor distance from the anal verge, and neoadjuvant therapy differed between with and without AL groups (P<0.2). These 3 variables were subsequently incorporated into the multivariate logistic regression model. The clinical model achieved an AUC of 0.766 in the training set and 0.583 in the validation set.

Radiomic feature extraction, selection, and development of the radiomics model
From T1- and fs-T2WI sequences, a total of 1,702 radiomic features were initially extracted. Of these, 1,356 features demonstrated ICCs between 0.75 and 1, and were therefore considered reproducible. These features were subsequently entered into the LASSO regression analysis to identify the most predictive characteristics. Nine features with nonzero coefficients were ultimately selected to construct the radiomics model. The optimal regularization parameter λ was determined as 0.0065 under the 1–standard error criterion (Fig. 3, Table 2). The Rad score was calculated for each patient using the aforementioned formula.

Development of the radiomics nomogram combined model and evaluation of the different models
The radiomics nomogram combined model was constructed by integrating Rad scores with independent clinical factors, including age, tumor distance from the anal verge, and preoperative neoadjuvant therapy. The regression coefficients and ORs for each variable were calculated (Supplementary Table 1). The ROC curves for each model in the training and validation sets are shown in Fig. 4. Performance metrics, including sensitivity and specificity, are summarized in Table 3 for both datasets. In both training and validation sets, the radiomics nomogram combined model demonstrated superior discriminative performance and higher AUC values compared with the clinical- and radiomics-only models. Decision curve analysis for the 3 models indicated that the radiomics nomogram combined model provided greater net clinical benefit across a wide range of threshold probabilities compared to either the clinical model or the radiomics model alone (Fig. 5).

DISCUSSION
Preoperative prediction of AL is crucial for optimizing surgical strategies and postoperative management in colorectal cancer. In this study, we developed a radiomics-based nomogram that integrates clinical risk factors with preoperative MRI-derived imaging features, providing a noninvasive tool for predicting postoperative AL. The radiomics nomogram model demonstrated strong predictive performance, with AUCs of 0.869 and 0.858 in the training and validation sets, respectively.
Many studies have reported associations between AL and various clinical factors. In our study, age, tumor distance from the anal verge, and preoperative neoadjuvant therapy emerged as predictors for constructing the multivariate logistic regression model. These findings are consistent with previously identified risk factors for AL [6, 21].
The relationship between neoadjuvant therapy and AL following colorectal cancer surgery has been extensively examined but remains controversial. Several large, high-impact studies suggest that neoadjuvant chemoradiotherapy (nCRT) does not significantly increase AL risk. For example, Bosset et al. [22] reported that preoperative chemoradiotherapy improved local control without increasing surgical complications, including AL. Similarly, Peeters et al. [23], in a landmark randomized trial, found no significant difference in AL rates between patients treated with nCRT and those undergoing surgery alone.
In contrast, an increasing body of high-quality evidence suggests a potential association between neoadjuvant therapy and AL, particularly in rectal cancer. A comprehensive meta-analysis by Yang et al. [24], which included 49 studies encompassing more than 23,000 patients, demonstrated that neoadjuvant radiotherapy significantly increased AL risk, with a pooled OR of 1.23 (95% CI, 1.07–1.41; P=0.004). This suggests that radiation-induced tissue changes may impair anastomotic healing. Likewise, a recent large-scale retrospective cohort study by Li et al. [25] found that nCRT was an independent risk factor for postoperative re-leakage after stoma reversal, with an OR of 4.07 (95% CI, 1.17–14.21; P=0.03). These findings underscore the need for individualized treatment strategies and careful perioperative planning in patients undergoing neoadjuvant therapy.
In our study, the leak rate in the neoadjuvant group was higher than in the non-neoadjuvant group (18.5% vs. 5.0%), aligning with the above evidence and suggesting that neoadjuvant therapy may contribute to AL risk under certain clinical circumstances. Variations in surgical technique, radiation dose, and patient-specific factors may further influence this risk. Prospective, multicenter trials with standardized protocols are warranted to optimize patient selection and perioperative strategies.
We also evaluated the role of protective ostomy in AL occurrence. Among patients with a diverting stoma, the AL rate was 4 of 23 (17.4%), compared with 7 of 123 (5.7%) in those without a stoma. The OR was 3.49 (95% CI, 0.93–13.07; P=0.073). However, this finding likely reflects surgical decision-making: patients considered at high risk were more often given a protective ostomy, rather than the ostomy itself increasing leak risk. This potential reverse causality should be taken into account when interpreting such results.
Previous studies have shown that while diverting stomas may not substantially reduce the incidence of AL, they can mitigate its clinical severity. For instance, Matthiessen et al. [26], in a randomized controlled trial, demonstrated that protective stomas significantly reduced the need for reoperation after low anterior resection. Similarly, a meta-analysis by Hüser et al. [27] concluded that although diverting stomas did not lower the overall incidence of AL, they were associated with decreased morbidity and reduced severity in patients who developed leaks. These findings emphasize the importance of protective ostomy in reducing the clinical impact of AL, even if it does not prevent its occurrence.
Surgical technique is another important factor influencing AL risk. In our cohort, higher AL incidence was observed among patients undergoing low anterior resection and those without a protective stoma, though not all comparisons reached statistical significance. Previous studies have shown that low anastomosis height and procedures such as total mesorectal excision are associated with increased AL risk due to compromised blood supply, greater anastomotic tension, and technical challenges [28, 29]. The absence of a diverting stoma has also been linked to higher rates of clinically significant leakage, suggesting a possible protective effect. Larger studies are needed to confirm these associations and further elucidate underlying mechanisms.
However, the clinical factors model alone did not achieve a high AUC for predicting AL. In our study, some clinical factors, such as sex, did not differ significantly between with and without AL groups, which contrasts with findings from previous reports [8, 30], This discrepancy may be attributable to the limited sample size and single-center design, underscoring the need for larger, multicenter studies.
Currently, no effective method exists for preoperative prediction of AL in colorectal cancer. Although a left-sided AL risk score has been developed, its external validation has produced unsatisfactory results [31]. Novel approaches such as intraoperative assessment of anastomotic perfusion may offer predictive value, but they also prolong operative time [32]. In contrast, our study provides a noninvasive and cost-effective method for preoperative AL prediction, representing an important clinical advantage.
Radiomics offers a promising tool for quantifying tumor heterogeneity and peritumoral tissue characteristics from routinely acquired images, detecting subtle features that are not discernible to the human eye. While previous radiomics research has largely focused on predicting tumor metastasis and treatment outcomes, less attention has been paid to surgical complications. Importantly, imaging features such as circumferential resection margin, extramural vascular invasion, and tumor deposits are key determinants in surgical planning [33, 34]. Certain factors, such as anastomotic perfusion and pelvic inflammation, have been confirmed to be associated with the occurrence of AL [35, 36]. Radiomics can capture and quantify these imaging characteristics, which may explain the strong predictive performance of our nomogram. Supporting this, previous radiomics-based models have also demonstrated predictive value for postoperative complications, such as pancreatic fistula [37]. Our findings therefore support the feasibility of incorporating radiomic biomarkers into predictive models for AL, providing added value beyond clinical variables alone.
In this study, we integrated clinical variables with MRI-derived radiomic features to develop a radiomics nomogram for AL prediction, achieving excellent predictive accuracy. To our knowledge, this represents one of the few studies employing an MRI-based radiomics model specifically for AL risk assessment in colorectal cancer.
Our study may have important clinical implications. Identification of high-risk patients preoperatively enables proactive measures such as intraoperative adjustments in surgical approach, placement of abdominal or anal drainage tubes, or the creation of a diverting stoma [38].
Our radiomics nomogram integrates MRI-derived radiomic features and key clinical factors to provide a personalized risk estimation for postoperative AL in colorectal cancer patients. In clinical practice, this tool can be used preoperatively to assist surgeons in stratifying patients according to their individual risk of developing AL. By inputting specific values of the predictors into the nomogram, clinicians can obtain a quantified probability of AL occurrence in an intuitive and user-friendly manner. This visualization facilitates rapid interpretation and supports risk stratification without requiring advanced computational tools.
Once high-risk patients are identified, tailored perioperative strategies may be implemented. These include conservative surgical approaches (e.g., temporary stoma formation), reinforcement of the anastomosis, application of sealants, or intensified postoperative surveillance (e.g., early imaging, closer clinical monitoring). Conversely, for patients with a low predicted risk, unnecessary protective interventions, such as diverting stomas that may impair quality of life, can be avoided. In this way, the nomogram enables more personalized and balanced decision-making, consistent with the principles of precision medicine.
The clinical value of our nomogram lies in supporting early risk prediction and informed surgical planning, with the potential to reduce both the incidence and clinical severity of AL. Given that AL is associated with increased morbidity, prolonged hospitalization, delayed initiation of adjuvant therapy, and greater healthcare costs, early identification of high-risk patients could mitigate these adverse outcomes. Additionally, the nomogram standardizes decision-making across clinicians by offering an evidence-based, reproducible tool. Its integration into preoperative workflows may enhance patient counseling, improve allocation of medical resources, and strengthen shared decision-making between surgeons and patients.

Limitations
Despite the promising findings, several limitations of this study must be acknowledged. First, the incidence of AL in our cohort was relatively low (11 out of 146 patients, 7.5%), which, although consistent with real-world clinical practice, significantly limits the statistical power for detecting meaningful associations and increases the risk of overfitting, even though we employed machine learning techniques to mitigate this issue. Moreover, the small number of leakage events may have constrained the identification of potential risk factors and reduced the robustness of model training. This also limits the generalizability of our findings to broader clinical populations.
Second, this study was conducted at a single center with a retrospective design, and internal validation was performed by randomly splitting the dataset into training and validation cohorts. Due to the lack of an independent external validation cohort, the generalizability and robustness of our model remain to be further verified. We acknowledge this important limitation and emphasize the need for future studies to include external or temporally distinct validation cohorts. Currently, we are actively collecting patient data from different time periods to facilitate such validation; however, the relatively low incidence of AL has slowed the accumulation of adequate sample sizes for meaningful analysis.
Third, radiomic features were extracted solely from noncontrast MRI sequences (T1WI and T2WI), which may restrict the imaging information captured. Incorporating contrast-enhanced or diffusion-weighted imaging could improve model comprehensiveness.
Lastly, variations in surgical techniques and the experience level of individual surgeons, both known risk factors for AL, were not fully accounted for due to data limitations.
To enhance the clinical applicability and translational potential of our model, future studies should focus on larger, prospective, multicenter cohorts with standardized imaging protocols and external validation across diverse institutions and populations. Integration of perioperative variables and real-time clinical data may also further optimize predictive performance.

Conclusions
Our study suggests that the proposed radiomics nomogram model may serve as a potentially useful tool for estimating the risk of postoperative AL in cancer patients. Moreover, transforming the nomogram into an automated software platform or online calculator could support its integration into clinical workflows and enhance decision-making in perioperative management.