Magnetic resonance imaging-based radiomics of mesorectum for predicting extramural venous invasion in patients with rectal cancer: a bi-centric study

Abstract

Objectives

To develop and validate a magnetic resonance imaging (MRI)-based radiomics model of the mesorectum for predicting extramural venous invasion (EMVI) in patients with rectal cancer (RC).

Methods

A retrospective study included 238 patients with RC from two hospitals between May 2020 and March 2023. Patients were divided into a training set (n = 114, from institution 1), an internal validation set (n = 48, from institution 1), and an external validation set (n = 76, from institution 2). A total of 963 radiomics features were extracted from the mesorectum region using T2-weighted imaging (T2WI). The radiomics model was developed using the methods of the minimum redundancy of the maximum relevance (mRMR) and the least absolute shrinkage (LASSO) regression. After univariate and multivariate logistic analysis, a clinical model was constructed based on clinical characteristics. A combined model was built and demonstrated as a nomogram. These models were evaluated by discrimination, calibration, and clinical application.

Results

Among 238 patients, 98 (41.1%) were EMVI-positive. The Area Under the Curve (AUC) values for the clinical, radiomics, and combined models, respectively, were 0.65, 0.85, and 0.88 for the training set (95% CI: 0.81–0.94); 0.60, 0.81, and 0.81 for the internal validation set (95% CI: 0.68–0.95); and 0.60, 0.78, and 0.82 for the external validation set (95% CI: 0.72–0.91).

Conclusion

This study presents a model for predicting the EMVI status in patients with rectal cancer. The combined model, which incorporates both a mesorectal radiomics signature from T2WI and the clinical predictor of serum neutrophil count, demonstrated superior discrimination, calibration, and clinical utility compared to models based on either clinical or radiomics features alone. The non-invasive tool shows promise for aiding in preoperative risk stratification and guiding clinical decision-making.

Key point

A combined model was developed for preoperative EMVI prediction using radiomics scores from T2W features and clinical information.

The combined model demonstrated superior diagnostic performance compared to either the radiomics model or clinical model alone for predicting EMVI in patients with RC.

Introduction

Rectal cancer (RC) ranks among the most prevalent malignancies of the digestive system, with an increasing incidence in younger demographic of patients [1]. Among its various biological characteristics, extramural venous invasion (EMVI) is recognized as a significant prognostic indicator and a key driver of local recurrence [2, 3]. Historically, EMVI was defined by Talbot et al. [4] as malignant cells within endothelium-lined vessels extending beyond the muscularis propria. In 2003, MRI-detected extramural vascular invasion (mrEMVI) was defined by Brown et al. [5] as “serpentine extension of tumor signal within a vascular structure.” Subsequently, Smith et al. [6] developed a standardized scoring system for assessing mrEMVI status, which was further popularized by the MERCURY study group as a reproducible and prognostically validated method for risk stratification in rectal cancer [7].

Typically, the gold standard for EMVI assessment is post-surgical pathology examination, which cannot help preoperative treatment decisions for patients. However, identifying EMVI status in smaller veins (< 3 mm) remains challenging due to standard MRI spatial resolution limitations. In recent years, radiomics has emerged as a non-invasive, convenient tool for extracting thousands of high-throughput quantitative features from digital images, offering a new dimension in personalized treatment [8, 9]. Previous studies [10–12] in RC have mostly focused on the intratumoral region, which would provide important but easily overlooked information on EMVI. Recently, many radiomics features resulting from the peritumoral area have also shown a good predictive power in response assessment in other carcinomas. For instance, Sun et al. [13] and Hu et al. [14] built radiomics models based on intratumoral and peritumoral radiomics, which showed a dependable ang high predictive ability of assessing the response to neoadjuvant chemoradiation in patients with cervical cancer and esophageal squamous cell carcinoma.

To our knowledge, there are no researches on the assessment of the status of EMVI by using radiomics of the entire mesorectum in patients with RC. In this study, we aim to construct and validate an MRI-based radiomics model of the mesorectum in a bi-centric database to predict the status of EMVI in patients with RC.

Materials and methods

Patients

This retrospective study enrolled consecutive patients with rectal cancer from Hunan Provincial People’s Hospital (institution 1) between May 2020 and October 2022 and Hunan Cancer Hospital (institution 2) between May 2022 and March 2023. Patients from institution 1 were used as the training and internal validation set, and those from institution 2 were used as the external validation set.

The inclusion criteria were as follows: (a) histopathological confirmed rectal cancer; (b) pelvic MRI examination performed within 2 months prior to surgery; (c) complete clinical data. The exclusion criteria were: (a) incomplete clinical data; (b) poor MRI quality prohibiting accurate mesorectal fascia segmentation; (c) any prior chemotherapy or other treatments before surgery. The details and baseline characteristics of the study are summarized in Table 1. Additionally, the process of patient selection flowchart was shown in Fig. 1.

Table 1.

Clinical characteristics in the training, internal validation, and external validation sets

Fig. 1.

Flowchart shows the patient selection from two hospitals in this study

Clinical characteristics and pathology information

Clinical characteristics, including sex, age, and preoperative laboratory test results such as carcinoembryonic antigen (CEA) level, carbohydrate antigen 19 − 9 (CA19-9) level, and neutrophil count, were collected from the hospital information system (HIS) database.

The gold standard for EMVI status was determined by postoperative histopathological evaluation. All surgically resected specimens were evaluated by pathologists who defined EMVI as a rounded mass of tumor tissue within an endothelium-lined space, which contained red blood cells or was surrounded by a rim of smooth muscle. In addition, a radiologist with 10 years of experience, who was blinded to the histopathologic and clinical information, retrospectively assessed the mrEMVI status on T2-weighted images using Smith’s 5-point scoring system. Pathological information also included lymph node metastasis and tumor stage according to the 8th edition of the American Joint Committee on Cancer (AJCC) tumor-node-metastasis (TNM) Staging System.

MRI protocol

MRI were performed with 4 MRI scanners from two institutions. For institution 1, patients were examined with 3.0-T (Magnetom Tim Trio, Siemens Healthineers), 1.5-T (Achieva, Philips Healthcare), and 3.0-T MRI systems (Ingenia, Philips Healthcare). For institution 2, patients were scanned on a 3.0-T MRI scanner (Discovery 750 W^®, GE Healthcare, Waukesha, WI). The axial T2-weighted imaging (T2WI) (perpendicular to the long axis of the rectum): TR/TE, 2700–5500/80–120; field of view (FOV), 180 × 180 to 240 × 240; matrix size, 320 × 320 to 580 × 365; slice thickness, 3 mm without an interslice gap. For hospital 2, the axial T2WI: TR/TE 3800–5500/80–120; FOV, 180 × 180 to 240 × 240; matrix size, 320 × 320; slice thickness, 3 mm without an interslice gap.

MRI radiomic analysis

Image segmentation

The axial T2WI was derived from the picture archiving and communication system (PACS). Two radiologists (with 3 and 5 years of experience in abdominal radiological diagnosis, respectively) manually segmented the entire mesorectum on each slice using 3D Slicer version 5.0.3 (www.slicer.org), an open-source and freely available software platform. The international consensus definition of the rectum is the “sigmoid take-off” which identifies the junction of the sigmoid mesocolon and the mesorectum, and therefore the sigmoid colon with the rectum [15]. The radiologists were blinded to histopathological results and clinical outcomes.

Furthermore, to evaluate inter-observer reproducibility, both radiologists delineated the volume of interest (VOI) on a randomly selected subset of 30 patients from the training set. The intraclass correlation coefficient (ICC) was calculated for each feature. Features with an ICC greater than 0.75 were considered reproducible and retained for further analysis [16].

Radiomic feature extraction

To ensure reproducibility of radiomic features across different scanners and acquisition protocols, standardized image preprocessing was performed prior to feature extraction. Specifically, all images were resampled to an isotropic voxel size of 1 × 1 × 1 mm using B-spline interpolation. Image intensity normalization was then conducted using z-score standardization.

The radiomics features of the mesorectum based on T2WI were extracted using Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/). After the mesorectum of the patients with rectal cancer was reconstructed and segmented, VOI images were processed through SlicerRadiomics code for feature extraction. All radiomics features of the VOI images were extracted, including: (1) first-order statistics, (2) 3D shape-based features, (3) texture features, and (4) wavelet features. In total, 963 radiomics features were extracted from each patient’s mesorectum region on T2WI.

Feature selection and model building

Patients from Institution 1 were randomly divided into a training cohort and an internal validation cohort at a 7:3 ratio, with the training cohort used for feature selection and model development. A two-step feature selection process was applied within the training cohort. First, mRMR was performed to eliminate redundant features, reducing the feature set to the top 20. Second, LASSO regression with 10-fold cross-validation was used to select the most informative and predictive features from the remaining 20. The final radiomics signature was constructed based on these selected features and their corresponding LASSO coefficients.

Statistical analyses

Continuous variables were compared using the Mann-Whitney U test or t test, and the categorical variables were compared using the χ2 test or Fisher’s exact test. Univariate and multivariate logistic regression analyses were used to identify independent predictors of EMVI. The area under the receiver operating curve (AUC) values of different models were compared by using the DeLong test. Decision curve analysis (DCA) was used to evaluate the clinical utility. And calibration curves indicated the goodness-of-fit of the combined model. The optimal cutoff values of sensitivity, specificity and accuracy values were acquired by the Youden index. The interobserver agreement between the two radiologists was analyzed by calculating the intraclass correlation coefficient (ICC). All statistical analyses were performed with R software (version 4.0.4; http://www.r-project.org/). p < 0.05 was considered statistically significant.

Results

Patient characteristics

A total of 238 consecutive patients (mean age, 63 ± 10 years; 153 male) were included in the study. From institution 1, 162 patients were included and randomly divided into a training set (n = 114) and an internal validation set (n = 48). The external validation set from institution 2 consisted of 76 patients. Overall, 98 of 238 patients (41.2%) were pathologically confirmed to be EMVI-positive. Significant differences were observed in T stage, N stage, and mrEMVI score between the EMVI-positive and EMVI-negative groups across all three sets (P < 0.05). In the training set, univariable analysis revealed that only mrEMVI score (P < 0.001) and serum neutrophil count (P < 0.05) were significantly associated with EMVI status. These were included in the multivariate analysis, which confirmed them as independent predictors. Detailed patient characteristics are presented in Table 1. In the training set, univariable analysis of variables revealed mrEMVI (P < 0.05) and serum neutrophils (P < 0.05) were associated with higher odds for patients of EMVI-positive and were thus included in the multivariable analysis. However, there was no statistically significant difference in other variables (P = 0.07 to P = 0.75) between the EMVI-negative and EMVI-positive groups (Table 2).

Table 2.

Logistic regression analysis of variables for the association with extramural venous invasion in patients in the training set

Models building and performance testing

Following mRMR and LASSO regression analysis, seven predictive radiomics features were selected to construct the radiomics signature (Fig. 2). The diagnostic performance of the clinical, radiomics, and combined models is detailed in Table 3 and Fig. 3. The clinical model, which included mrEMVI score and serum neutrophil count, yielded AUCs of 0.65, 0.60, and 0.60 in the training, internal validation, and external validation sets, respectively. In comparison, the radiomics model showed significantly better performance, with corresponding AUCs of 0.85, 0.81, and 0.78. The combined model, which integrated the radiomics signature with the two clinical predictors, demonstrated the highest discrimination capability, with an AUC of 0.88 (95% CI: 0.81, 0.94) in the training set, 0.81 (95% CI: 0.68, 0.95) in the internal validation set, and 0.82 (95% CI: 0.72, 0.91) in the external validation set (Fig. 3).

Fig. 2.

The workflow of the construction and internal verification of a radiomics model

Table 3.

Diagnostic performance of the clinical, radiomics, and combined models in the testing set, internal validation, and external validation sets

Fig. 3.

Assessment of models for the ability to predict extramural venous invasion status. Areas under the receiver operating characteristic curve (AUCs) for the (A) training set (n = 114), (B) internal validation set (n = 48), and (C) external validation set (n = 76). AUCs are reported with 95% CI in parentheses

The DeLong test showed that the combined and radiomics models were both significantly superior to the clinical model (P < 0.05). There was no statistically significant difference between the combined and radiomics models (P > 0.05), although the combined model showed a consistently higher net benefit in DCA. A nomogram was developed based on the combined model to provide an individualized tool for predicting EMVI probability (Fig. 4). The calibration curves demonstrated good agreement between the nomogram’s predictions and actual observations in all three sets (Fig. 5A). DCA showed that the combined model provided a greater net benefit across a wide range of threshold probabilities compared to either the radiomics or clinical model alone, indicating superior clinical usefulness (Fig. 5B).

Fig. 4.

Nomogram of the combined model for predicting extramural venous invasion status (EMVI) in patients with rectal cancer

Fig. 5.

(A) The calibration curves of the training, internal validation, and external validation sets indicated the goodness-of-fit of the clinical-radiomic nomogram. The dotted line represents the performance of the nomogram in different sets and a closer fit to the diagonal solid line represents a better prediction. (B) Decision curve analysis of these models to investigate the clinical usefulness in predicting EMVI. The blue line represents the decision curve for the radiomics model. The red line represents the clinical model and the green line represents the combined model. Using the combined model to predict the EMVI status added more benefit than using either the treat-all plan or the treat-none plan at any given threshold probability

Smith’s 5-point scoring system performance

The diagnostic performance of the Smith 5-point scoring system was evaluated using postoperative pathological findings as the reference standard (Table 4). Overall, the scoring system showed moderate diagnostic accuracy for EMVI assessment. Among the individual score categories, patients classified as score 1 achieved the highest accuracy (83%), whereas score 4 was associated with a lower accuracy (46%) and a higher rate of misclassification compared with other score levels. In addition, the Smith scoring system showed a tendency to overestimate EMVI status, with a higher proportion of patients assigned positive EMVI scores than confirmed by pathology.

Table 4.

Diagnostic performance of the smith 5-point scoring system and the combined model in the training and validation cohorts

Discussion

In this study, we developed and validated an MRI-based radiomics model using features from the mesorectum to predict EMVI status in patients with RC. Our main finding is that the combined model, which integrated the mesorectal radiomics signature with clinical risk factors, achieved the best predictive performance, with AUCs of 0.88, 0.81, and 0.82 in the training, internal validation, and external validation sets, respectively. A nomogram incorporating the radiomics signature and clinical predictors was also developed to provide a more precise method for predicting EMVI status. Although some overlap was observed in the confidence intervals of the AUCs, the DeLong test indicated a statistically significant difference between models. More importantly, DCA demonstrated that the combined model provided greater net benefit across clinically relevant thresholds, supporting its potential clinical value beyond conventional statistical metrics.

A key finding of this study was that an elevated serum neutrophil count was an independent predictor of EMVI. Quantitatively, multivariate analysis demonstrated that increased neutrophil levels were associated with a significantly higher risk of EMVI (OR = 2.73, 95% CI: 1.16–6.44, P < 0.05). This observation is consistent with prior studies on inflammation-based prognostic scores, such as the neutrophil-to-lymphocyte ratio, in which neutrophils represent a key inflammatory component [17, 18]. From a biological perspective, neutrophils may contribute to the development of EMVI through the formation of neutrophil extracellular traps, which can disrupt vascular endothelial integrity and facilitate tumor cell adhesion and intravascular dissemination [19]. Neutrophil extracellular traps (NETs), released during a process called NETosis, can disrupt endothelial cell junctions, compromise vascular integrity, and subsequently mobilize tumor cells, potentially leading to venous thromboembolism.

Preoperative assessment of EMVI is crucial for clinical decision-making, as it is a known independent risk factor for poor survival and recurrence in RC. However, the traditional diagnosis relies on postoperative surgical specimens, which cannot guide preoperative therapy. Some previous studies have used different MRI sequences to predict EMVI status, Bae et al. [20] showed an inexperienced reviewer detected the status of EMVI on MRI using a 5-point grading system with poor performance (AUC of 0.658). Crimì et al. [21] detected EMVI in patients with rectal cancer using T2-weighted, diffusion weighted imaging (DWI) and contrast-enhanced sequences, DWI (AUC of 0.729) showed better performance for EMVI detection compared to contrast-enhanced (AUC of 0.624) and T2-weighted sequences (AUC of 0.610).

Unlike previous studies [22, 23] that primarily extracted radiomic features from the gross tumor volume, our study specifically focused on the entire mesorectum. From a biological perspective, the mesorectum represents the primary microenvironment for rectal cancer progression and local spread. Radiomic features derived from this region may reflect subtle alterations in the peritumoral microenvironment, including perivascular inflammatory changes and stromal remodeling, which are known to precede overt extramural vascular invasion detectable on conventional imaging. By incorporating these peritumoral characteristics, our approach may capture complementary information beyond intratumoral heterogeneity alone, thereby contributing to the improved predictive performance observed in our model.

In recent years, some researchers have found that radiomics provide a noninvasive and reproducible method to elicit undiscovered features from imaging data that are not accessible by radiologist perception or conventional image analysis [24]. Shu et al. [25] reported that the Bayes-based radiomics signature predicted EMVI with an AUC of 0.744. Yu et al. [26] also showed that the dynamic contrast-enhanced magnetic resonance imaging-based radiomics (AUC, 0.812) had good prediction of assessing EMVI. These studies [27] can assess the status of EMVI by MRI-based radiomics feature, to create a standardized definition and normalized reference for radiomics research, the image biomarker standardization initiative criteria and radiomics quality score were established [28]. However, these studies only focused on the associations between tumor imaging and radiomics features to predict EMVI.

Recently, a few studies [13, 14, 29] indicated that the surrounding areas can provide additional information about tumor heterogeneity in other cancers such as esophageal squamous cell carcinoma and cervical cancer. Actually, the mesorectum is the place where EMVI occurs. Therefore, we proposed a widely available, noninvasive, and MRI-based radiomics model with a favorable predictive value using mesorectum radiomics features to predict the status of EMVI in patients with rectal cancer. The strong performance of our radiomics-only model (AUCs of 0.85, 0.81, and 0.78 across the three sets) supports this hypothesis. To our knowledge, this is the first study to investigate the association between mesorectal radiomic features and EMVI status. In addition to comparative performance, the generalizability of radiomics models remains a key concern for clinical translation [30]. In the present study, acceptable predictive performance was maintained in an independent external cohort acquired using a different scanner platform, suggesting a certain degree of robustness. Nevertheless, scanner heterogeneity and the lack of fully standardized acquisition protocols may still influence radiomic feature stability. Although intensity normalization and resampling were applied, we acknowledge that more advanced harmonization techniques and prospective protocol standardization were not implemented. Future multicenter studies incorporating standardized imaging protocols, formal harmonization methods, and validation across diverse scanner platforms are therefore warranted to further improve reproducibility and generalizability.

This study has several limitations. First, its retrospective design is susceptible to inherent selection bias. Second, correlating EMVI on a preoperative MRI with the exact location in the postoperative surgical specimen can be challenging. Third, the manual segmentation of the mesorectum was time-consuming and labor-intensive; fully automated segmentation using artificial intelligence is a goal for future work. Fourth, the data from multiple centers were acquired using different MRI scanners, which may influence radiomic features. Finally, the study was limited to a single external validation cohort due to data availability.

In conclusion, this study establishes that a model combining clinical data with radiomic features from the mesorectum can accurately predict EMVI status in patients with RC. This noninvasive approach may serve as a promising imaging marker to enhance preoperative risk stratification and guide clinical practice. Further validation in larger, prospective studies is required to confirm its clinical utility.

Conclusion

In summary, this study presents a model for predicting the EMVI status in patients with rectal cancer. The non-invasive tool shows promise for aiding in preoperative risk stratification and guiding clinical decision-making.

Author contributions

Peng Liu and Xiaoping Yu conceived of and designed the study and polished the language of the manuscript; Jia He wrote the initial draft of the paper; Jia He, Xiang Feng, Jiabei Liu and Diejuan Liu collected and analysed the data; Jia He, Huashan Lin and Jing Fang conducted the statistical analysis; All authors approved of the final version before submission.

Funding

This study have received by the Changsha Municipal Natural Science Fundation [grant numbers: kq2014201] and the Youth Medical Key Talents Foundation of Hunan Province [grant no. 20230508-1034].

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethical approval and consent to participate

Ethics license was received from the Science Ethics Committee. This study is a retrospective study and informed consents were waived.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.