Rectal cancer (RC) is one of the most prevalent malignant tumors of the digestive tract. It ranks as the third leading cause of cancer-related mortality worldwide (1). A substantial proportion of patients present with locally advanced RC (LARC) at initial diagnosis (2). According to the National Comprehensive Cancer Network guidelines (3), the standard treatment paradigm for LARC includes neoadjuvant chemoradiotherapy followed by total mesorectal excision (TME) after a 5–12-week interval, with optional postoperative adjuvant chemotherapy. However, a recent multicenter randomized trial demonstrated that a neoadjuvant chemotherapy-only regimen in patients with LARC at a relatively low risk of recurrence (T2N1/2, T3N0/N1) is not inferior to preoperative chemoradiotherapy in terms of disease-free survival and local recurrence rates (4). This finding underscores the clinical imperative of accurately identifying T3N0/N1 RC preoperatively, as it enables tailored, risk-adapted treatment strategies and potentially mitigates unnecessary treatment-related toxicity.

Magnetic resonance imaging (MRI) serves as the preferred method for RC staging, achieving exceptional T-stage classification accuracy of 88–99% through high-resolution soft tissue characterization (5). Nevertheless, its diagnostic performance in classifying lymph node metastasis (LNM) remains suboptimal (accuracy: <80%) and is constrained by its reliance on size-morphology criteria with inherent limitations (6). Pathological analyses revealed that 28% of metastatic nodes measure ≤3 mm in short-axis diameter (7), fundamentally challenging conventional size thresholds. While functional MRI techniques, particularly diffusion-weighted imaging (DWI), show potential for improved nodal characterization through cellularity assessment via apparent diffusion coefficient (ADC) mapping (8), their diagnostic performance remains moderate (accuracy: 66%, sensitivity: 53%, and specificity: 82%), Significant ADC value overlap between reactive and malignant nodes necessitates complementary diagnostic strategies (9).

Radiomics enables the extraction of high-dimensional quantitative features from medical images. It has emerged as a powerful tool for overcoming conventional imaging limitations (10) and may serve as a valuable adjunct in assessing LNM in RC (11). However, most existing studies have primarily focused on presence vs. absence of LNM (12), with limited exploration of predictive models in terms of specific lymph node staging (e.g., N0/N1 vs. N2). The present study aimed to develop and validate a radiomics model based on high-resolution T2-weighted MRI to differentiate between N0/N1 and N2 stages in stage T3 RC patients using multiple machine learning algorithms. It also sought to establish the clinical utility of radiomics-based nodal staging as a noninvasive diagnostic tool by systematically evaluating the predictive performance of various models. A more precise lymph node staging framework may enable risk-adapted treatment strategies, such as chemotherapy-based regimens for low-risk patients (e.g., T3N0/N1), reduce unnecessary radiotherapy exposure, minimize the risk of overtreatment, and ultimately improve patient outcomes and quality of life.

The lymph node status in RC is a critical factor in determining the necessity of adjuvant therapy and surgical resection (18). However, Initial MRI radiology report for lymph node (LN) staging in rectal cancer relies on subjective size-morphology criteria, often showing poor interobserver agreement (κ = 0.416) and limited accuracy (AUC: 0.60–0.62), remains challenging (19). In contrast, Dong et al. (20)demonstrated that a radiomics model achieved higher diagnostic precision (PPV: 75.9%) by quantifying tumor heterogeneity, overcoming the limitations of traditional MRI. Building upon the recently published novel perspectives and conclusions (4), an exploratory investigation was conducted in the present study. Five radiomics models were developed based on high-resolution T2WI features using five different machine learning algorithms in order to noninvasively differentiate between N0/N1 and N2 stages in stage T3 RC. The performance and clinical utility of these models were systematically compared. The results showed that the LR model exhibited the best performance in predicting lymph node staging in stage T3 RC. The clinical applicability of this model was further validated via calibration curves and DCA, indicating positive clinical benefits at most threshold probabilities.

Radiomics features can reveal subtle changes in rectal tumor lesions that are difficult to discern with the naked eye (21, 22). In the present study, 1,746 radiomics features were extracted from oblique axial T2W images. LASSO regression was subsequently utilized to select 16 key features for in-depth analysis. These included one first-order and 15 texture features, with texture features being the most abundant, accounting for 93.75% of the total. This result is consistent with those described by Li and Yin et al. (23), who demonstrated that T2WI-based texture features have high accuracy in predicting LNM in RC (AUC of 0.805), further validating the importance of texture features in tumor assessment.

GLSZM features were predominant in the present study. GLSZM quantifies the size distribution of continuous image regions, finely characterizing the distribution patterns of homogeneous regions of different sizes within tumors. Smaller regions may correspond to densely packed tumor cells or micro-nodules, while larger continuous regions may reflect necrotic areas or fibrotic changes. (24). This spatial heterogeneity is closely related to tumor proliferation activity and invasiveness. Studies have shown that the short-zone emphasis in GLSZM features is significantly negatively correlated with tumor microvascular density, while large-zone emphasis is positively correlated with stromal fibrosis (25). This ability to quantify spatial heterogeneity in tumor microstructures makes GLSZM features an important indicator for assessing tumor invasiveness and metastatic potential. Therefore, texture features can capture the microscopic heterogeneity within tumors, reflecting the spatial arrangement and gray-level distribution patterns of tumor cells (26–28). These features showed significant differences in distinguishing between N0/N1 and N2 stages (p < 0.05) in the present study, indicating that LNM is closely related to microscopic structural changes within the tumor. Additionally, the predictive model based on T2WI radiomics features was constructed using wavelet features (8/16). Wavelet transform is a multi-scale analysis method with perfect reconstruction capability, ensuring no information loss or redundancy during signal decomposition. It can decompose images into high-frequency (heterogeneity) and low-frequency (homogeneity) components, facilitating the extraction of structural information and details from the original images (29). Previous studies (30) have also reported the effectiveness of wavelet features in predicting lymph node status in T2WI. Furthermore, He et al. (31) found that wavelet features in T2WI performed well in RC tumor grading, further demonstrating that they can represent the biological behavior and heterogeneity of tumors.

A growing body of evidence underscores the value of clinical parameters as complementary predictors in radiomics-based models, with integrated frameworks often demonstrating superior discrimination compared with radiomics alone (12). For example, (32) reported that a combined clinical–radiomic model improved sensitivity (82.6% vs. 78.3%) and specificity (88.9% vs. 57.9%) relative to a purely radiomic approach, while also mitigating the subjectivity of conventional MRI interpretation through visualized risk maps. Such clinical metrics may encode systemic or tumor-related biological processes that radiomics, which predominantly captures spatial and textural heterogeneity, cannot fully represent. In our cohort, however, none of the routinely collected clinical variables demonstrated statistically significant intergroup differences across outcome categories (all P > 0.05). This absence of discriminative signal suggests that, in this specific population, these variables may have limited incremental value for the prediction task. From a modeling perspective, the inclusion of non-informative covariates in high-dimensional feature spaces risks diluting true signal, inflating model variance, and impairing generalizability—particularly in datasets of modest size. Given these considerations, and in pursuit of parsimony, we elected to exclude clinical parameters from the final model. Several factors may underlie this discrepancy with prior studies. First, differences in patient demographics, disease stage distribution, and treatment patterns between our cohort and other trials could attenuate the predictive contribution of clinical variables. Second, sample size constraints may have limited statistical power to detect subtle effects, especially for variables with low intergroup variability. Third, the endpoints examined—derived from imaging-based nodal staging—may be more tightly coupled to local morphologic features than to systemic clinical markers. These hypotheses warrant systematic evaluation in larger, multi-institutional datasets, ideally with harmonized variable definitions and broader biological characterization, to clarify the true translational potential of integrating clinical and radiomic predictors.

The widespread application of machine learning algorithms in the field of radiomics has significantly improved diagnostic performance (33). Selecting the appropriate classifier is crucial for building high-performance predictive models. The present study systematically evaluated five supervised learning models commonly used for binary classification tasks, including LR, SVM, Bernoulli Naive Bayes, ridge regression, and SGD. The results showed that the LR model performed the best, with AUCs of 0.900 and 0.876 in the training and testing sets, respectively. Additionally, the model demonstrated excellent accuracy (training set: 0.847, testing set: 0.882) and specificity (training set: 0.873, testing set: 0.927). The superior performance of the LR model could be attributed to its ability to effectively handle linearly separable data, low complexity after feature selection, and resistance to overfitting. Moreover, the LR model has strong interpretability, allowing for the quantification of each feature’s contribution to the prediction results (34), which provides important references for clinical decision-making. Similarly, Wei et al. (35) developed and validated a clinical radiomics model by combining T2W and amide proton transfer-weighted MRI radiomics features, achieving efficient LNM prediction in rectal adenocarcinoma using an LR classifier (AUCs of 0.983, 0.864, and 0.851 in the training, validation, and testing sets, respectively). Cui et al. (36) used an LR model to predict a complete pathological response in LARC and achieved an AUC of 0.90. However, the LR model is not suitable for all scenarios, as classifier performance highly depends on the distribution characteristics of the training and testing sets. Qu et al. (37) found that the SVM classifier performed best when constructing a predictive model using T2W images, with AUCs of 0.892 and 0.71 in the training and validation sets, respectively. This is because SVM can handle nonlinear relationships through kernel functions and exhibits strong classification capabilities (38).

This study has several limitations. First, this study focused on evaluating MRI-based radiomics for lymph node staging in patients undergoing upfront surgical resection for stage T3 rectal adenocarcinoma. However, select patients with low-risk, distally located tumors (e.g., T3N0/N1) achieving a clinical complete response after neoadjuvant therapy may be eligible for a watch-and-wait (W&W) organ-preserving strategy. Excluding such cases may limit the model’s applicability in settings where nonoperative management is considered. Second, the model was developed solely on T2-weighted MRI data, potentially limiting sensitivity for detecting small or morphologically subtle lesions. Future work should explore the integration of functional imaging sequences—such as diffusion-weighted imaging or contrast-enhanced MRI-within a multiparametric framework to enhance lesion conspicuity and improve diagnostic accuracy. Third, manual ROI delineation inevitably introduces interobserver and intraobserver variability, which may bias feature extraction and subsequent model performance. The adoption of automated or semi-automated segmentation algorithms could reduce subjectivity, standardize feature generation, and improve reproducibility across centers. Finally, the modest sample size and absence of external validation restrict the model’s generalizability. Rigorous validation using large, multicenter datasets with diverse patient populations is essential to confirm robustness, refine calibration, and establish the clinical utility of the proposed model.

The present study confirmed the effectiveness of a machine learning-based high-resolution T2WI radiomics model in predicting lymph node staging in stage T3 RC. Radiomics is a noninvasive assessment method that can provide a valuable alternative for lymph node staging, supporting personalized treatment decisions. It showed broad application prospects in optimizing treatment pathways, avoiding overtreatment, and improving the prognosis of patients with LARC. Future research may further advance the clinical application of this technology via multimodal imaging and large-scale validation.