Artificial intelligence-integrated multimodal data-assisted magnetic resonance imaging for neoadjuvant chemoradiotherapy decision-making in cT1-2N0 rectal cancer

Abstract

BACKGROUND

Colorectal cancer represents a major global health burden, with neoadjuvant chemoradiotherapy as standard treatment for locally advanced rectal cancer, although preoperative magnetic resonance imaging (MRI) staging shows only moderate accuracy. A substantial proportion of patients staged as cT1-2N0 on MRI are subsequently upstaged after surgery, thereby missing neoadjuvant treatment and risking worse oncologic outcomes.

AIM

To develop a machine learning model that integrates multiomics profiles to improve the accuracy of neoadjuvant chemoradiotherapy decision-making in rectal cancer patients whose baseline MRI indicates cT1-2N0 disease.

METHODS

This multicenter cohort study consecutively enrolled patients who underwent pre-operative MRI and curative rectal cancer surgery at three institutions between January 2013 and December 2024. Participants were randomly allocated to training, internal validation and external validation sets. A pre-defined set of 2260 radiomic features was extracted from T2-weighted and diffusion-weighted images and fused with baseline clinical, hematological and pathological variables. Ten independent machine learning classifiers were constructed and compared with both physician decisions and imaging-only radiomic models. Model performance was evaluated with receiver operating characteristic analysis, calibration plots, decision curve analysis and confusion matrices.

RESULTS

A total of 1320 consecutive patients with clinically staged cT1-2N0 rectal cancer who underwent curative intent surgery were retrospectively enrolled from three tertiary centers: (1) Center 1 (n = 1009); (2) Center 2 (n = 246); and (3) Center 3 (n = 65). We developed a Clinical, Hematologic, Oncopathologic, and Radiomic Decision (CHORD) model, an XGBoost-based machine learning classifier that integrates seven preoperative MRI radiomic features with 16 clinical, hematologic, and pathologic variables via a CART regression tree algorithm in patients with cT1-2N0 rectal cancer. Externally validated, the CHORD model delivered an area under the curve of 0.927 and an F1-score of 0.825, attesting to its consistent and robust performance across independent cohorts.

CONCLUSION

Our study demonstrated that the CHORD model exhibits satisfactory performance in identifying cT1-2N0 rectal cancer patients who require neoadjuvant chemoradiotherapy by integrating preoperative multi-omics data, offering critical insights into reducing the omission rate of neoadjuvant therapy and enhancing the accuracy of preoperative radiological staging.

Key Words: cT1-2N0 rectal cancer; Magnetic resonance imaging; Machine learning; Radiomics; SHapley Additive exPlanations; Surgery

Core Tip: We developed an artificial intelligence-driven multimodal model that not only improves radiomic performance but also provides clinicians with a reliable tool to reduce missed identification of patients who would benefit from neoadjuvant therapy. Compared with radiologist assessment, the model demonstrates improved reliability and clear advantages in supporting clinical decision-making.

INTRODUCTION

Colorectal cancer imposes a substantial global health burden, and its incidence and mortality have doubled in dozens of regions worldwide[1,2]. Neoadjuvant chemoradiotherapy is the standard of care for patients with defined cT3-4N0 or node-positive rectal cancer; it not only downsizes the primary tumor but also achieves pathological complete response in a subset of cases, thereby enabling R0 resection and improving long-term oncologic outcomes[3-5]. However, patients with cT1-2N0 rectal cancer are primarily treated by radical surgery alone[6,7]. In clinical decision-making for neoadjuvant chemoradiotherapy in rectal cancer patients, a genuine dilemma exists: Preoperative magnetic resonance imaging (MRI) remains the cornerstone for staging, yet its accuracy is only moderate, ranging from 60% to 78%[8].

Standalone preoperative MRI staging of rectal cancer remains suboptimal; a non-negligible subset of patients radiologically classified as cT1-2N0 are ultimately upstaged to pT3-4 or node-positive disease after surgery[9,10]. This systematic understaging frequently leads to omission of neoadjuvant chemoradiotherapy and may consequently compromise oncologic outcomes[11,12]. Consequently, stage-missed neoadjuvant candidates (SNC) staged as cT1-2N0 by preoperative MRI pose a therapeutic dilemma: Occult disease, imaging mimicry, and the need to refine early-stage classification to enable deployment of potentially curative neoadjuvant therapy underscore the urgent demand for accurate, efficient decision-support tools.

Radiomics non-invasively transforms conventional clinical images into high-dimensional quantitative data, enabling dissection of tumor biology and serving as a surrogate for pivotal genomic processes underlying rectal carcinogenesis[13-15]. Prior studies have demonstrated that MRI-derived imaging signatures can predict pathological complete response after neoadjuvant therapy in rectal cancer, thereby enabling a select cohort of patients to adopt a watch-and-wait strategy and informing tailored treatment decisions[16,17]. Machine learning – an artificial intelligence-based paradigm – holds considerable promise for maximizing the clinical utility of radiomics[18,19]. Artificial intelligence-driven, multi-modal imaging prediction models have demonstrated superior performance, illuminating a viable path forward for future clinical decision making[20,21].

Accordingly, this study sought to develop a pre-operative model that integrates clinical variables with MRI-derived radiomic features to predict the need for neoadjuvant chemoradiotherapy in SNC staged as cT1-2N0 rectal cancer, thereby refining the accuracy of pre-operative imaging stratification. The predictive capacity of the model was subsequently evaluated and externally validated in two independent cohorts from separate institutions.

MATERIALS AND METHODS

Study design and population

A total of 1009 patients with pathologically confirmed rectal cancer who underwent radical resection at The First Affiliated Hospital of Air Force Medical University (Center 1) between January 2013 and December 2024 were retrospectively analyzed. Center 2 comprised patients treated at the Chinese PLA General Hospital from January 2016 to December 2024, and Center 3 comprised rectal cancer patients treated at Shaanxi Provincial People’s Hospital from January 2022 to December 2024. With its more recent data, Center 3 provided a better foundation for application of the model.

Inclusion criteria: (1) Histologically confirmed primary rectal adenocarcinoma; (2) Clinical stage cT1-2N0 on contrast-enhanced pelvic MRI; and (3) No neoadjuvant chemotherapy or radiotherapy prior to surgery.

Exclusion criteria: (1) Missing essential clinical variables; (2) Incomplete MRI data or severe imaging artifacts; (3) Distant metastasis; and (4) Microsatellite instability-high (MSI-H) or DNA mismatch repair (dMMR) status determined by immunohistochemistry.

To ensure robust evaluation, Center 1 data were randomly split 7:3 into training and internal validation sets, while patients from Centers 2 and 3 were pooled to form an independent external validation cohort, all enrolled under the same inclusion and exclusion criteria applied to Center 1 (Figure 1).

Figure 1 Patient selection flowchart. dMMR: DNA mismatch repair; MRI: Magnetic resonance imaging; MSI-H: Microsatellite instability-high.

Clinical data collection

Preoperative baseline characteristics – including age, sex, hematological indices (e.g., red cell count) and pathological data (e.g., D2-40 staining) – were retrieved from the institutional electronic medical record system. Features with fewer than 20% missing values were retained, and missing values were imputed using multiple imputation (Supplementary Tables 1 and 2, Supplementary Figure 1). All pelvic MRI studies were independently reviewed by two board certified radiologists and two gastrointestinal surgeons to assign clinical T and N stage; discordant cases were adjudicated by the senior staff radiologists of the center (Supplementary Table 3). All resected specimens were precisely pathological staged according to the TNM classification of the 8^th edition of the American Joint Committee on Cancer Staging Manual[22].

Image data preprocessing

MRI images for Center 1 were acquired on a GE Discovery MR750 3.0-T (GE Healthcare) superconducting scanner using an eight-channel abdominal phased array coil. The acquisition parameters were as follows: For axial T1-weighted fast spin echo sequences, the repetition time (TR) was set between 3.8 milliseconds and 700 milliseconds, the echo time (TE) between 1.7 milliseconds and 7.8 milliseconds, the image matrix was 288 pixels × 224 pixels to 320 pixels × 224 pixels, the slice thickness was 4-7 mm, and the field of view (FOV) was 38 cm × 38 cm to 42 cm × 42 cm. For T2-weighted sequences, TR ranged from 2300 milliseconds to 5119 milliseconds, TE from 84.1 milliseconds to 102.5 milliseconds, the matrix size was 288 pixels × 224 pixels to 320 pixels × 224 pixels, the slice thickness was 6-7 mm, and FOV was 38 cm × 38 cm to 44 cm × 44 cm. For diffusion-weighted imaging (DWI) with a b value of 1000, TR was 4800-5000 milliseconds, TE was 59.2-60 milliseconds, the matrix was 128 pixels × 128 pixels to 160 pixels × 160 pixels, the slice thickness was 6-7 mm, and FOV was 36 cm × 36 cm to 44 cm × 44 cm. These parameters collectively ensured the quality and accuracy of MRI examinations for rectal cancer.

MRI images for Center 2 were acquired on a GE Discovery MR750 3.0 T (GE Healthcare) superconducting scanner using an eight-channel abdominal phased array coil. The acquisition parameters were as follows: For axial T1-weighted fast spin echo sequences, TR was set between 530 milliseconds and 700 milliseconds, TE was 11.1 milliseconds, the image matrix was 320 pixels × 256 pixels, the slice thickness was 5 mm, and the FOV was 38 cm × 38 cm. For T2-weighted sequences, TR ranged from 2700 milliseconds to 4939 milliseconds, TE from 58 milliseconds to 100 milliseconds, the matrix size was 288 pixels × 224 pixels to 320 pixels × 224 pixels, the slice thickness was 4-7 mm, and FOV was 24 cm × 20 cm to 42 cm × 42 cm. For DWI with a b value of 800, TR was 3000-6650 milliseconds, TE was 62.9-65 milliseconds, the matrix was 128 pixels × 128 pixels to 160 pixels × 160 pixels, the slice thickness was 4-7 mm, and FOV was 24 cm × 20 cm to 44 cm × 44 cm. MRI images for Center 3 were acquired using Signa HDxt 1.5 T (GE Healthcare) and Discovery 750w 3.0 T (GE Healthcare) scanners.

MRI images for Center 3 were acquired using Signa HDxt 1.5 T (GE Healthcare) and Discovery 750w 3.0 T (GE Healthcare) scanners. The acquisition parameters were as follows: For axial T1-weighted imaging, TR was 150-630 milliseconds, TE was 1.5-14 milliseconds, matrix was 256 × 230 to 320×272, slice thickness was 5-7 mm, and FOV was 19.1 cm × 12.7 cm to 28 cm × 19.6 cm. For axial T2-weighted imaging, TR was 2000-4970 milliseconds, TE was 83-131 milliseconds, matrix was 256 × 179 to 288 × 288, slice thickness was 4-5 mm, and FOV was 15 cm × 10.5 cm to 40 cm × 40 cm. For DWI with a b value of 1000, TR was 3689 milliseconds, TE was 75.4 milliseconds, matrix was 112 × 114, slice thickness was 5 mm, and FOV was 11.2 cm × 11.4 cm. These parameters collectively ensured the quality and accuracy of MRI examinations for rectal cancer.

Prior to any analysis, all images were harmonized through intensity normalization and grayscale standardization to eliminate brightness and contrast variability across scanners; every subsequent radiomic feature was extracted exclusively from these normalized datasets (Supplementary Table 4).

Tumor lesion segmentation and feature extraction

To ensure unbiased assessment, two surgeons (Kang BY, 3 years of experience; Qiao YH, 6 years of experience) manually delineated slice by slice tumor regions of interests (ROIs) on T2-weighted and DWI images while blinded to all patient information; the contours were subsequently verified by a senior colorectal radiology surgeon with > 10 years of expertise.

PyRadiomics (v3.0.1) was employed to extract radiomic features from the T2-weighted (T2W) and DWI images within the manually delineated ROIs[23]. A total of 1130 radiomic features were extracted per ROI and sequence. These comprised 216 first-order statistics (including 18 based on original images, 144 based on wavelet decompositions, and 54 based on log-sigma LOG filters), 900 texture features (including 75 based on original images, 600 based on wavelet decompositions, and 225 based on log-sigma LOG filters) distributed across five matrix classes – gray-level co-occurrence matrix, gray-level dependence matrix, gray-level run length matrix, gray-level size zone matrix, and neighboring gray tone difference matrix – and 14 shape descriptors.

Radiologist diagnosis

In the conventional workflow, preoperative staging was based on the radiologists’ visual assessment. For the present comparison, the same 1130 patients were independently reviewed by two intermediate radiologists without the aid of the proposed model; discordant cases were adjudicated by a senior consultant radiologist, whose diagnosis served as the reference clinical standard. Native DICOM images were loaded into RadiAnt DICOM Viewer, which allowed the physicians to measure maximal lesion diameter. From admission onward, patients were managed by gastrointestinal surgeons; radiologists – aware only of imaging, sex and age – assigned preoperative T and N stages, and the final T/N categories recorded after senior attending review were used for analysis.

Model development

To mitigate scanner related radiomic feature drift across centers, inter-site concordance was first quantified and the efficacy of Z-score harmonization was verified by t-SNE dimensionality reduction. A total of 2260 radiomic features (1130 each from T2W and DWI) were extracted using PyRadiomics (Supplementary Figure 2). Multivariable logistic regression prefiltering retained 176 features with P > 0.05, from which tenfold least absolute shrinkage and selection operator (LASSO) regression selected 7 informative variables (three T2W derived, four DWI derived). Clinical variables were extracted from electronic medical records, and the training and validation sets were assigned by complete randomization. Features were selected exclusively from the training set; only those simultaneously deemed important by multivariable logistic regression, LASSO and Boruta random forest were retained, yielding 16 final covariates: (1) Two baseline clinical features; (2) Twelve hematologic indices; and (3) Two pathologic markers. Multicollinearity was assessed by variance inflation factor and Pearson correlation; variables with variance inflation factor > 5 or |r| > 0.8 were removed. Following feature selection, ten radiomics-only models and ten multimodal models integrating radiomics with clinical variables were trained. Hyperparameter optimization was performed with tenfold cross-validation; early stopping within each fold protected against overfitting. The best performing algorithm was retained for final analysis and designated the Clinical, Hematologic, Oncopathologic, and Radiomic Decision (CHORD) model.

Model evaluation

Three models were constructed: (1) Radiologists’ unassisted diagnoses; (2) A radiomics-only machine learning classifier; and (3) The multimodal CHORD predictor (Figure 2). The performance of each was then examined in two independent external validation cohorts. Receiver operating characteristic curves were plotted and area under the curve (AUC) values were computed; accuracy, sensitivity, specificity and F1 score were quantified. Calibration curves assessed model fit, and decision curve analysis evaluated clinical utility across probability thresholds. SHapley Additive exPlanations (SHAP) explainability was finally used to rank feature importance and to guide neoadjuvant chemoradiotherapy decisions at both population and individual patient levels. Each patient’s individual probability was compared with the population-level SHAP-derived risk threshold to identify high risk individuals likely to be upstaged postoperatively, thereby guiding neoadjuvant chemoradiotherapy decisions (Figure 2).

Figure 2 Schematic of Clinical, Hematologic, Oncopathologic, and Radiomic Decision machine-learning model construction and validation. A-C: They delineate the high-throughput radiomics feature extraction pipeline embedded within the Clinical, Hematologic, Oncopathologic, and Radiomic Decision framework; D: It illustrates the rigorous preprocessing of multidimensional clinical variables; E: It summarizes the comparative benchmarking of candidate models and the subsequent SHapley Additive exPlanations-driven post-hoc interpretability analysis. AUC: Area under the curve; DWI: Diffusion weighted imaging; GBM: Gradient boosting machine; KNN: K-nearest neighbors; LASSO: Least absolute shrinkage and selection operator; ROC: Receiver operating characteristic; ROI: Regions of interest; SVM: Support vector machine; 3D: Three-dimensional.

Statistical analysis

Statistical analysis was performed in R, version 4.4.1. Radiomic features were extracted with 3D Slicer (v5.8.1) using the SlicerRadiomics extension (commit 8426cdf) and with PyRadiomics (v3.0.1) under Python (v3.13.7). LASSO regularization model, Boruta regression, and Logistic regression were calculated using the R packages Boruta and glmnet. The ten machine learning models were generated using the following R packages: E1071, gradient boosting machine, caret, XGBoost, NNET, Adaboost, LightGBM and CatBoost. Decision curve analysis and calibration curves were calculated using the R packages pROC, RMDA and riskRegression. All continuous variables are displayed as mean ± SE, and all categorical variables are displayed as frequencies and percentages. The χ² test was used to verify the differences between groups, and P < 0.05 was considered statistically significant. The kernelshap and shapviz packages in R were used to ascertain the significance and hierarchy of variables within the model.

Ethics approval

The study protocol adhered to the ethical guidelines of the 1995 Declaration of Helsinki, and this study was approved by the ethics committee of the First Affiliated Hospital of Air Force Medical University (No. KY20232232-C-1). The study was based on research data registered on the World Health Organization International Clinical Trials Registry Platform (ChiCTR2300070629) on April 18, 2023. Given the retrospective nature of the cohort and the anonymization of all data, individual informed consent was waived. All patient data were fully de-identified. This study employed a retrospective design and strictly adhered to the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis + artificial intelligence statement, as well as the step-by-step guidance for developing clinical prediction models.

RESULTS

Patient baseline characteristics

Center 1 enrolled 1009 patients (573 males, 436 females; median age 61.0 years, range 22.0-88.0) (Table 1). This cohort comprised 756 non-SNC and 253 SNC cases. Centers 2 and 3 together contributed 267 non-stage missed and 44 SNC. Supplementary Table 1 summarizes the baseline characteristics of patients from each center.

Table 1 Baseline characteristic of patients, n (%)/median (interquartile range).

Variable	Non-SNC (n = 756)	SNC (n = 253)	P value
T stage			< 0.0011
1	89 (12)	80 (32)
2	667 (88)	173 (68)
Sex			< 0.0011
Female	303 (40)	133 (53)
Male	453 (60)	120 (47)
Fecal occult blood			0.241
Negative	173 (23)	49 (19)
Positive	583 (77)	204 (81)
Smoking and alcohol history			0.131
Negative	585 (77)	200 (79)
Smoking history	84 (11)	35 (14)
Alcohol consumption history	15 (2.0)	5 (2.0)
Combined smoking-alcohol history	72 (9.5)	13 (5.1)
Surgery history			0.0561
Negative	82 (11)	17 (6.7)
Positive	674 (89)	236 (93)
Cardiovascular disease			0.831
Negative	737 (97)	246 (97)
Positive	19 (2.5)	7 (2.8)
Hypertension			0.881
Negative	267 (35)	88 (35)
Positive	489 (65)	165 (65)
Diabetes			0.901
Negative	403 (53)	136 (54)
Positive	353 (47)	117 (46)
ABO blood group			0.471
A	326 (43)	109 (43)
B	186 (25)	55 (22)
AB	82 (11)	36 (14)
O	162 (21)	53 (21)
RH blood group			0.992
Negative	5 (1)	2 (1)
Positive	751 (99)	251 (99)
Histological type			< 0.0011
Ulcerative	517 (68)	133 (53)
Infiltrating	9 (1.2)	17 (6.7)
Protruding	230 (30)	103 (41)
Differentiation			< 0.0012
Poorly differentiated	114 (15.1)	15 (6)
Moderately differentiated	570 (75.4)	222 (88)
Well differentiated	72 (9.5)	15 (6)
Margin			> 0.992
Negative	6 (0.8)	2 (0.8)
Positive	750 (99)	251 (99)
S100			< 0.0011
Negative	106 (14)	74 (29)
Positive	650 (86)	179 (71)
CD34			> 0.992
Negative	1 (0.1)	0 (0)
Positive	755 (100)	253 (100)
D2-40			< 0.0011
Negative	192 (25)	157 (62)
Positive	564 (75)	96 (38)
NRS2002			< 0.0012
0	151 (20)	61 (24)
1	333 (44)	95 (38)
2	158 (21)	59 (23)
3	53 (7.0)	36 (14)
4	61 (8.1)	0 (0)
5	0 (0)	2 (0.8)
Caprini			< 0.0012
0	8 (1.1)	2 (0.8)
1	13 (1.7)	31 (12)
2	243 (32)	84 (33)
3	304 (40)	78 (31)
4	145 (19)	43 (17)
5	35 (4.6)	10 (4.0)
6	6 (0.8)	2 (0.8)
7	2 (0.3)	3 (1.2)
Eastern Cooperative Oncology Group			< 0.0012
0	509 (67)	225 (89)
1	223 (29)	26 (10)
2	23 (3.0)	2 (0.8)
3	1 (0.1)	0 (0)
Age	61.000 (53.500-67.500)	62.000 (57.000-69.000)	0.0123
Body mass index	22.676 (20.950-25.000)	24.030 (23.000-26.600)	< 0.0013
Alfa-fetoprotein	2.795 (2.235-3.960)	3.130 (2.530-3.870)	0.0243
Carcinoembryonic antigen	2.260 (1.440-3.720)	2.280 (1.370-3.380)	0.293
CA19-9	10.650 (7.175-15.700)	11.000 (6.460-15.530)	0.703
CA724	3.210 (1.110-10.070)	2.510 (0.970-10.070)	0.0343
CA125	10.700 (8.250-14.600)	11.300 (8.520-15.900)	0.163
Ki67	0.750 (0.600-0.800)	0.800 (0.600-0.800)	0.0163

¹Pearson's χ² test.

²Fisher's exact test.

³Wilcoxon rank sum test.

CA: Carbohydrate antigen; SNC: Stage missed neoadjuvant candidates.

We found that the most statistically significant differences between SNC and non-SNC patients were in T stage, sex, NRS2002 score, Caprini score, Eastern Cooperative Oncology Group (ECOG) performance status, histological type, differentiation, S100 expression, and D2-40 expression (all P < 0.001). Specifically, the SNC group had a higher proportion of T1-stage tumors (P < 0.001) and a significantly higher proportion of female patients compared with male patients (P < 0.001). Additionally, the proportion of patients with an NRS20002 score ≥ 5 was significantly lower in the SNC group than in the non-SNC group (P < 0.001; Supplementary Tables 5-8).

Performance of different models

Variable selection was performed using LASSO regression and multicollinearity testing. Among the MRI radiomics features, the following seven were identified as predictive of SNC: (1) Three-dimensional surface area on T2W imaging (T2WI); (2) Shannon entropy of voxel intensities after 1.0 mm Gaussian filtering (first-order texture) on T2WI; (3) Gray-level co-occurrence matrix-based correlation coefficient at 1.5 mm smoothing scale on T2WI; (4) Gray-level variance within the low-high-low wavelet sub-band on DWI; (5) High-Gray-Level Run Emphasis in the high-high-low wavelet sub-band on DWI; (6) Variance of run lengths at 1.5 mm smoothing scale on DWI; and (7) Neighborhood gray-tone difference matrix contrast at 1.0 mm smoothing on DWI (Supplementary Table 9, Supplementary Figure 3). Clinical features deemed significant by logistic regression, Boruta random forest, and LASSO regression were selected for model construction (Supplementary Figures 4-6). Two multicollinearity tests confirmed that clinical features, radiomics features, and the combined feature set exhibited no multicollinearity (Supplementary Figure 7). A radiomics-based machine-learning model constructed using these seven features demonstrated good performance in predicting SNC. XGBoost was the optimal classifier, with an AUC of 0.677, accuracy of 0.611, and F1-score of 0.571. The two junior attendings and the senior radiologist achieved accuracies of 0.66, 0.75 and 0.79, sensitivities of 0.76, 0.84 and 0.92, and specificities of 0.63, 0.72 and 0.75, respectively (Supplementary Table 10). Although the radiomics machine-learning model underperformed relative to the clinicians, it offered markedly faster decision-making and more stable performance across cases. Among the radiomics-clinical fusion pipelines, the XGBoost-based model was selected as the final CHORD predictor. It demonstrated excellent discrimination for SNC: (1) Internal-validation AUC of 0.938, accuracy of 0.870, F1-score of 0.869; and (2) External-validation AUC of 0.927, accuracy of 0.920, F1-score of 0.825 (Table 2, Supplementary Table 6, Supplementary Figure 8).

Table 2 Performance comparison of ten machine-learning models trained on multimodal radiomics and clinical data vs radiologist diagnosis.

Model	Area under the curve	Accuracy	Sensitivity	Specificity	Positive predictive value	Negative predictive value	F1
Logistic	0.798 (0.739-0.857)	0.817	0.547	0.907	0.661	0.858	0.699
Support vector machine	0.817 (0.764-0.870)	0.744	0.827	0.717	0.492	0.926	0.717
Gradient boosting machine	0.917 (0.881-0.953)	0.904	0.760	0.951	0.838	0.923	0.897
NeuralNetwork	0.798 (0.739-0.857)	0.731	0.733	0.730	0.474	0.892	0.676
RandomForest	0.922 (0.889-0.954)	0.857	0.827	0.867	0.674	0.938	0.843
XGBoost	0.938 (0.911-0.965)	0.870	0.867	0.872	0.692	0.952	0.869
K-nearest neighbors	0.910 (0.874-0.946)	0.787	0.920	0.743	0.543	0.966	0.783
Adaboost	0.747 (0.684-0.810)	0.748	0.680	0.770	0.496	0.879	0.673
LightGBM	0.922 (0.889-0.956)	0.834	0.893	0.814	0.615	0.958	0.728
CatBoost	0.886 (0.845-0.928)	0.807	0.813	0.805	0.581	0.929	0.678
Reader 1	1	0.660	0.760	0.626	0.404	0.923	0.528
Reader 2	1	0.750	0.840	0.720	0.500	0.931	0.627
Reader 3	1	0.790	0.920	0.746	0.548	0.966	0.687

¹The physicians’ diagnostic decisions were recorded as binary categorical judgments (e.g., presence vs absence of disease) rather than as continuous probability scores. Therefore, receiver operating characteristic curve analysis and the corresponding area under the curve were not applicable. Instead, sensitivity and specificity were calculated.

Model performance evaluation

Model performance was comprehensively evaluated using receiver operating characteristic curves, calibration plots, decision-curve analysis, and confusion matrices (Figure 3). CHORD exhibited consistent performance across the training, internal-validation, and external-validation cohorts, underscoring its robustness and clinical utility (Figure 3A, E, and I). Although the standalone radiomics machine-learning model achieved acceptable discrimination, its calibration plot revealed a non-negligible deviation in the mid-probability range, and the decision curve showed only modest net clinical benefit between thresholds of 0.25 and 0.80. In contrast, the CHORD model – assessed in both internal and external cohorts – exhibited high accuracy on confusion-matrix analysis (Figure 3B, F, and J), well-calibrated predictions across the entire probability spectrum (Figure 3C, G, and K), and substantial net benefit for any threshold between 0.25 and 0.95, underscoring its superior and stable clinical utility (Figure 3D, H, and L).

Figure 3 Performance evaluation of the radiomics machine-learning model in internal validation and of the Clinical, Hematologic, Oncopathologic, and Radiomic Decision model in both internal and external validation. A-C: They display receiver operating characteristic curves; D-F: They show confusion matrices; G-I: They present calibration plots; J-L: Thet depict decision-curve analyses. AUC: Area under the curve; GBM: Gradient boosting machine; KNN: K-nearest neighbors; ROC: Receiver operating characteristic; SVM: Support vector machine.

SHAP-based model interpretation

To demystify the “black box” nature of artificial intelligence–driven machine learning models, SHAP-based algorithms were used to explicate the models’ internal logic and confer interpretability[24]. We aggregated cohort-level model features and quantified global feature importance using SHAP bar and swarm plots. We identified serum triglycerides, serum potassium, and ECOG performance score as the most important predictors of SNC (Figure 4A and B) and visualized individual feature importance with waterfall and force plots. SHAP evaluation of two SNC cases and one non-SNC case showed that serum triglycerides and serum potassium contributed the most, whereas D-40, ECOG performance score, and tumor histotype were secondary contributors (Figure 4C-E). Partial dependence plots illustrated variable interactions: When ECOG score was 1-2 or serum potassium was 4.3-5.5 mmol/L, serum triglyceride levels were typically below 5.6 mmol/L, and these two factors appeared to interact as joint protective determinants against SNC, yielding high combined feature contribution (Figure 4F and G, Supplementary Figures 9-11).

Figure 4 SHapley Additive exPlanations interpretation of the XGBoost model. A: SHapley Additive exPlanations (SHAP) summary bar plot; B: SHAP summary bee-swarm plot; C: SHAP force plot for individual explanation; D and E: SHAP waterfall plots for individual explanation; F: SHAP partial dependence plot for serum triglycerides and Eastern Cooperative Oncology Group performance score; G: SHAP partial dependence plot for serum triglycerides and serum potassium. ALT: Alanine aminotransferase; ECOG: Eastern Cooperative Oncology Group; SHAP: SHapley Additive exPlanations.

Based on the CHORD model, we selected patient 112 for individual analysis using a random number table. The patient was a 70-year-old male with rectal adenocarcinoma, clinically staged as cT2N0 and pathologically staged as pT2N1. Preoperative imaging revealed thickening of the rectal wall and intermediate-to-high signal intensity on T2W images (Figure 5A-D). Standardized digital rectal examination and pathological biopsy confirmed the diagnosis of cT2N0 rectal adenocarcinoma, with no neoadjuvant chemoradiotherapy administered. The preoperative MRI T2W and DWI-weighted images of the patient were segmented for ROI, with feature extraction performed for each slice (Figure 5E-H), followed by evaluation using the web-deployed CHORD model (Figure 5I). The CHORD model classified patient 112 as high-risk, with an SNC probability of 0.665 and a decision probability for neoadjuvant chemoradiotherapy of 0.9051 (Figure 5I and J). SHAP analysis revealed that serum triglycerides, serum potassium, and preoperative albumin contributed most to the auxiliary MRI diagnosis, with contributions of 0.1360, 0.1010, and 0.0900, respectively (Figure 5K). However, the patient experienced local recurrence and liver metastasis within 2 months after surgery, resulting in a poor prognosis. We also applied CHORD to a patient initially classified as non-SNC: The model assigned a probability of 0.426, consistent with the non-SNC label. The patient was discharged in April 2017, returned with recurrence in November 2022, and died in March 2023, yielding a PFS of 67 months and an OS of 71 months – indicating an overall favorable prognosis (Supplementary Figure 12).

Figure 5 Individual prediction and analysis of patient No. 112 using the Clinical, Hematologic, Oncopathologic, and Radiomic Decision Model. A: Axial section of the patient’s T2-weighted (T2W) magnetic resonance imaging (MRI) image; B: Coronal section of the patient’s T2W MRI image; C: Sagittal section of the patient’s T2W MRI image; D: The patient’s diffusion weighted imaging MRI image; E: Axial section of the patient’s T2W MRI image with radiologist’s segmentation; F: Coronal section of the patient’s T2W MRI image with radiologist’s segmentation; G: Sagittal section of the patients T2W MRI image with radiologist’s segmentation; H: The patient’s diffusion weighted imaging MRI image with radiologist’s segmentation; I: Web deployment of the Clinical, Hematologic, Oncopathologic, and Radiomic Decision model and individual prediction results; J: Bar chart of decisions made by Clinical, Hematologic, Oncopathologic, and Radiomic Decision and other machine-learning models; K: Bar chart of SHapley Additive exPlanations contributions of the patient’s features. CHORD: Clinical, Hematologic, Oncopathologic, and Radiomic Decision; ECOG: Eastern Cooperative Oncology Group; GBM: Gradient boosting machine; KNN: K-nearest neighbors; SHAP: SHapley Additive exPlanations; SVM: Support vector machine.

DISCUSSION

In this study, we developed a multimodal machine learning model based on radiomics and other modalities (model CHORD) to predict stage-missed candidates for neoadjuvant therapy in patients with T1-2N0 rectal cancer. We compared the CHORD model with nine other machine learning models and diagnoses by radiologists to assess its efficacy. The results showed that the physician’s decision making was superior to most machine learning models. The CHORD model, which integrates multimodal features, demonstrated the best ability to predict SNC and make decisions regarding neoadjuvant chemoradiotherapy.

Patients with SNC were older and more likely to be female than those without SNC; however, neither age nor sex emerged as significant risk factors for SNC. Our study demonstrated that SNC is characterized by an earlier T stage and lower NRS2002 score, Caprini score, and ECOG performance status. These patients presented with early T stage disease, excellent performance status, and imaging findings lacking overt high-risk features; consequently, subtle microinvasion was consistently underrecognized, and the final pathological stage was routinely upstaged following surgery. Pathologically, SNC differed significantly from non-SNC in S100 and D2-40 expression (P < 0.001). Consistent with prior reports, S100-negative and D2-40-negative staining were characteristic of SNC[25,26]. Our study highlights the performance gain achieved by integrating radiomics with clinical data. This unified modeling strategy anchored the algorithm to real world practice and offered the prospect of higher diagnostic accuracy and immediate clinical utility.

Neoadjuvant chemoradiotherapy is the standard of care for cT3-4N0 or node-positive rectal cancer; it cuts locoregional relapse, prolongs survival, and converts borderline unresectable tumors to clear R0 resections[27-29]. However, preoperative imaging remains inherently limited: With T and N stage accuracy stabilizing around 72%, a substantial fraction of rectal cancers is assigned cT1-2N0, only to be reclassified as pT3–4N0 or N+ after resection[30]. Given the irreversibility of the omission of neoadjuvant chemoradiotherapy that ensues, refining the precision of baseline radiologic assessment is imperative. Moreover, the noninvasive and readily accessible nature of imaging confers decisive advantages for large-scale screening and first-line detection[31,32]. To increase preoperative MRI accuracy in cT1-2N0 rectal cancers with minimal tumor burden and to precisely predict SNC, we developed the CHORD model. CHORD integrates non-invasive radiomic signatures extracted from standard high-resolution MRI with readily available clinical variables, routine hematological indices, and mandatory pretreatment biopsy data. Deploying an ensemble machine learning pipeline, the platform augments baseline MRI performance and furnishes an individualized recommendation for neoadjuvant chemoradiotherapy in patients initially staged as cT1-2N0.

In the internal validation cohort and two independent external datasets, CHORD consistently delivered robust discrimination and calibration, translating its predictive precision into clinically actionable stratification that directs neoadjuvant chemoradiotherapy toward cT1-2N0 patients most likely to benefit. SHAP analysis clarified several candidate radiomic, hematologic and histopathologic features that most plausibly predict synchronous nodal commitment. Serum triglyceride, potassium and ECOG performance score emerged as the top three contributors to CHORD’s probability of synchronous nodal commitment; low triglyceride, low potassium and excellent ECOG status were consistently flagged as risk factors for occult nodal disease. Hypotriglyceridemia provided the single largest SHAP contribution to SNC prediction, reflecting tumor-driven metabolic reprogramming, declining systemic inflammatory nutritional reserve, and occult tumor burden. Preoperative hypotriglyceridemia has been associated with a 42% postoperative stage shift in locally advanced rectal cancer, presumably owing to the greater occult tumor burden present in neoplasms that rely on fatty acid oxidation[33]. This biology driven shift supported the notion that early nutritional support or neoadjuvant chemoradiotherapy may confer disproportionate benefit in patients with low serum triglyceride levels. Conversely, the blunted systemic inflammatory response and occult micrometastatic foci associated with low triglyceride levels are both linked to postoperative upstaging[34,35]. Prior basic science studies have demonstrated that serum potassium itself does not directly predict SNC; instead, it acts as a readily accessible surrogate marker for systemic electrophysiological-metabolic-immune imbalance[36-38]. Mechanistically, hypokalemia slows intestinal smooth muscle motility and weakens the gut immune barrier while simultaneously promoting angio- and perineural invasion through the transforming growth factor-beta/Smad axis, processes that often evade MRI detection and together can drive synchronous nodal commitment[39,40]. A low ECOG performance score emerged as a risk factor for synchronous nodal disease; we posit that this paradox reflects both a blunted systemic inflammatory response consistent with prior data and clinician underestimation of early danger signals in seemingly fit patients[41].

The CHORD model delivers early detection of SNC and generates patient-specific recommendations for neoadjuvant chemoradiotherapy. After standard staging MRI is completed for all cT1-2N0 patients, the images are automatically forwarded to the clinician-facing CHORD interface, which returns an immediate SNC risk probability and treatment suggestions. Moreover, CHORD matched board-certified radiologists in overall accuracy and outperformed clinician judgment when guiding decisions on neoadjuvant chemoradiotherapy, while cutting diagnostic turnaround time by nearly half. Early clinical use showed that patients selected by CHORD for neoadjuvant chemoradiotherapy, originally staged as cT1-2N0, later yielded mesorectal or internal iliac nodes that were occult at baseline MRI yet sterilized in the resection specimen, underscoring the therapeutic value of model-guided treatment for SNC. However, patients whom the model flagged as SNC-positive, but who declined immediate neoadjuvant chemoradiotherapy, still appeared to benefit after choosing nutritional support or additional PET-CT surveillance, suggesting that early identification alone can guide beneficial alternatives even when standard treatment is deferred.

Several limitations of this study should be acknowledged. First, the exclusion of incomplete or low-quality images may have introduced selection bias. Second, the comparison with radiologists was restricted to one intermediate-level and one senior reader from our own institution; other centers and additional senior physicians were not included. Full inclusion of radiologists at all experience levels and from additional centers will be incorporated into future trials. Third, we have not yet examined model stability, long-term accuracy, or the potential risk of overtreatment – issues that require dedicated surveillance. Finally, inadequate assessment of inter-site imaging variability and the exclusion of dMMR/MSI-H patients who received neoadjuvant immunotherapy restrict generalizability. Ongoing work will therefore develop a dedicated machine-learning model for the dMMR/MSI-H population and will incorporate multi-center, multinational cohorts, with prospective validation planned through a multicenter clinical trial.

CONCLUSION

In summary, the individualized multimodal CHORD classifier demonstrated robust discrimination and treatment selection value for synchronous nodal commitment, offering cT1-2N0 patients a reference standard and potentially curative therapeutic path, which may provide an accurate, efficient, and cost effective tool for clinical practice.

ACKNOWLEDGEMENTS

We thank the First Affiliated Hospital of Air Force Medical University for its support, the Department of Clinical Epidemiology and Health Statistics of Air Force Medical University for statistical guidance.