Deep learning radiomics nomogram based on multi-regional features for predicting lymph node metastasis and prognosis in colorectal cancer

Abstract

BACKGROUND

Lymph node metastasis (LNM) is closely linked to poor prognosis in patients with colorectal cancer (CRC). An accurate preoperative evaluation of lymph node status is crucial for tailoring individualized treatment plans. However, dependable methods for predicting lymph node involvement and assessing prognosis prior to surgery are still inadequate.

AIM

To develop a nomogram combining clinical features, radiomics and deep learning (DL) features for predicting LNM and prognosis in patients with resectable CRC.

METHODS

Two hundred and seventy-eight patients with pathologically confirmed CRC from two hospitals were retrospectively enrolled. Radiomics and DL features were extracted from preoperative three-phase contrast-enhanced computed tomography images within the intratumoral and peritumoral-3 mm regions. Using logistic regression combined with three feature selection methods, a clinical-DL-radiomics nomogram (DLRN) was developed. SHapley Additive exPlanations were employed to interpret the model. Additionally, Cox regression analysis was performed to identify risk factors for 3-year recurrence-free survival and to construct a prognostic model.

RESULTS

The DLRN demonstrated significantly superior performance in differentiating LNM compared to both the clinical model and the radiomics model. The area under the receiver operating characteristic curve for the DLRN was 0.944, 0.878, and 0.855 in the training, internal validation, and external validation cohorts, respectively. Calibration curves and decision curve analysis confirmed that the DLRN exhibited excellent calibration and strong clinical utility. Furthermore, the prognostic model based on the DLRN score showed robust performance in predicting 3-year recurrence-free survival, achieving an area under the receiver operating characteristic curves of 0.826, 0.788, and 0.755 across the three cohorts, respectively.

CONCLUSION

The DLRN demonstrated excellent performance in predicting LNM in CRC. Additionally, the prognostic model derived from the DLRN score effectively stratified patients according to their risk of recurrence.

Key Words: Colorectal cancer; Radiomics; Deep learning; Lymph node metastasis; Prognosis; Contrast-enhanced computed tomography; Nomogram; SHapley Additive exPlanation

Core Tip: Accurate preoperative prediction of lymph node metastasis is crucial for optimizing treatment strategies in colorectal cancer. In this study, we developed an interpretable clinical-deep learning-radiomics nomogram (DLRN) by integrating clinical features with multi-regional radiomics and deep learning features. Moreover, the DLRN-based prognostic model effectively predicted 3-year recurrence-free survival. As a noninvasive preoperative tool, the DLRN demonstrated strong predictive accuracy for lymph node metastasis in colorectal cancer and offers a practical means for individualized risk stratification and informed treatment decision-making.

INTRODUCTION

Colorectal cancer (CRC) is one of the most prevalent malignant tumors encountered in clinical practice. According to the 2022 Global Cancer Statistics Report, CRC ranks as the third most common malignancy worldwide and is the second leading cause of cancer-related mortality[1]. Lymph node metastasis (LNM) represents the most frequent metastatic route in CRC and is a critical factor contributing to postoperative recurrence and patient mortality[2]. Importantly, the presence or absence of LNM not only directly influences the extent of surgical intervention and the formulation of adjuvant therapy regimens but also significantly affects long-term patient survival[3,4]. Therefore, accurate preoperative assessment of lymph node status holds considerable clinical significance for optimizing personalized treatment strategies and prognostic evaluation in CRC patients.

Currently, contrast-enhanced computed tomography (CECT) is a commonly used method for the preoperative assessment of LNM in CRC. However, its accuracy in detecting LNM is limited by several factors, including equipment performance, the standardization of imaging protocols, the expertise of radiologists, and individual patient variability. Radiomics, a rapidly evolving field, shows promising potential for LNM prediction by extracting high-throughput quantitative features from medical images and analyzing them[5,6]. Concurrently, deep learning (DL), a key branch of machine learning, can automatically extract and characterize deep-level tumor image features in an end-to-end manner[7], achieving remarkable results in tumor diagnosis and prognostic prediction[8-10]. Several studies have demonstrated that although radiomics and DL individually possess considerable classification capabilities, they each excel at capturing different aspects of image information, radiomics in texture and morphology, and DL in high-level semantic features, and their combination often yields superior performance[11,12]. Nevertheless, current research on LNM prediction in CRC primarily focuses on the independent use of either radiomics or DL features, with limited exploration of their effective integration.

CRC exhibits substantial spatial heterogeneity, encompassing both the tumor itself (intratumoral region) and the surrounding peritumoral area. Although peritumoral tissue may appear similar to normal tissue macroscopically, studies have revealed significant microstructural heterogeneity within this region[13]. Moreover, tumor cells have a biological propensity to infiltrate and migrate into the peritumoral area, suggesting that this region may harbor critical information related to LNM[14,15]. With the advancement of radiomics, analysis of peritumoral features has become increasingly prevalent in tumor prediction research and has demonstrated considerable value in LNM prediction[16,17]. For instance, Yuan et al[18] reported the potential of computed tomography-based intratumoral and peritumoral radiomics models for preoperative LNM prediction in rectal cancer; however, their study lacked a comprehensive external validation cohort, limiting its clinical applicability.

Therefore, this study aims to develop and validate a comprehensive nomogram [DL-radiomics nomogram (DLRN)] that integrates clinical features with multi-regional radiomics and DL features for the preoperative prediction of LNM in CRC patients. Furthermore, building upon this predictive model, a prognostic model for 3-year recurrence-free survival (RFS) is constructed to provide a reliable tool for individualized clinical decision-making and risk stratification.

MATERIALS AND METHODS

Study population

This retrospective study was approved by our institutional review board, with a waiver of informed consent, approval No. 2025-158-01. Between January 1, 2018, and December 31, 2021, a total of 412 patients with resectable CRC were consecutively enrolled from the First People’s Hospital of Hefei (Center 1) and Binhu Hospital of Hefei (Center 2). Inclusion criteria were: (1) Pathologically confirmed CRC following surgery; (2) CECT performed within two weeks prior to surgery; and (3) No history of other malignant tumors. Exclusion criteria included: (1) Poor CECT image quality due to bias, motion artifacts, or low resolution; (2) Receipt of neoadjuvant chemoradiotherapy or other treatments before surgery; and (3) Incomplete clinical or pathological data. Ultimately, 220 patients from Center 1 and 58 patients from Center 2 met the criteria and were included. The 220 patients from Center 1 were randomly divided into a training cohort (n = 154) and an internal validation cohort (n = 66) at a 7:3 ratio, while the 58 patients from Center 2 formed the external validation cohort. The study population flowchart is presented in Figure 1.

Figure 1 Flowchart of patient enrollment. CRC: Colorectal Cancer; CECT: Contrast-enhanced computed tomography; LNM: Lymph node metastasis.

Image acquisition and evaluation

Arterial, venous, and delayed phase CECT images were retrieved from the picture archiving and communication system for further analysis. Detailed CT acquisition parameters for each center are provided in Table 1. A non-ionic iodinated contrast agent (300-400 mgI/mL) was administered intravenously at a dose of 1-1.5 mL/kg, with an injection rate of 3 mL/second. After contrast administration, contrast images were obtained at the arterial phase (25-35 seconds), portal venous phase (65-80 seconds), and delayed phase (3-5 minutes). To minimize batch effects arising from different scanners, ComBat harmonization was applied. Additionally, to standardize spatial resolution and intensity across datasets, all original CECT images were resampled to an isotropic voxel size of 1 mm × 1 mm × 0.625 mm (slice thickness 0.625 mm), and z-score normalization was performed on image intensities to reduce scanner-related variability. Two experienced radiologists independently evaluated the CECT images: Radiologist A, with 5 years of experience in CRC imaging, and radiologist B, with 12 years of experience. Both were blinded to clinical and pathological information. In cases of disagreement, a senior radiologist (radiologist C), with over 20 years of expertise in CRC CECT diagnosis, reviewed the conflicting cases and facilitated a consensus decision.

Table 1 Computed tomography protocols of the two centers.

Center	Scanner	Field of view (cm)	Image matrix (mm)	Tube voltage (kv)	Tube current (mA)	Pixel size (mm)	Slice thickness (mm)
1	Revolution CT (256-slice)	50	512 × 512	70-140	Auto	0.977	0.625-5
	Discovery CT750 HD (Gemstone)	50	512 × 512	80/100/120/140	Auto	0.977	0.625-5
2	SOMATOM definition flash (dual-source CT)	50	512 × 512	70-140	Auto	0.977	0.625-5

CT: Computed tomography.

Image segmentation

The radiomics workflow is illustrated in Figure 2. Radiologist A manually segmented the three-dimensional (3D) region of interest of the tumor using ITK-SNAP software (version 3.8.0; www.itksnap.org). To capture information from the invasive margin, the peritumoral region was automatically generated by expanding the tumor boundary outward by 3 mm using Python-based morphological erosion and dilation algorithms, resulting in a 3 mm-thick ring around the tumor. When this ring overlapped with adjacent structures such as the bladder, kidney, or surrounding normal intestine, these areas were manually excluded. To assess the reproducibility of radiomics and DL feature extraction, radiologist A and radiologist B randomly selected 50 patients for repeated region of interest segmentation two weeks later, including both the intratumoral region and the automatically defined 3-mm peritumoral region. Intra-class and inter-class correlation coefficients (ICCs) were then calculated for the features derived from both regions to evaluate intra-observer and inter-observer reliability.

Figure 2 Workflow of the study. ROI: Region of interest; DL: Deep learning; ROC: Receiver operating characteristic; CA 19-9: Carbohydrate antigen 19-9; CEA: Carcinoembryonic antigen; DLRR: Deep learning-radiomics integrated model; SHAP: SHapley Additive exPlanation.

Radiomics model construction

Radiomics feature extraction was conducted using the Pyradiomics package (version 2.2.0) in Python (version 3.7.1) on the CECT images. Radiomics features were extracted from each volume of interest and included first-order statistics, gray-level co-occurrence matrix, gray-level run length matrix, gray-level size zone matrix, gray-level dependence matrix, and shape features. Based on the training cohort data, the radiomics features were filtered and reduced: (1) All radiomics features were standardized using z-score normalization to eliminate scale differences; (2) Reproducibility was assessed using the ICC, and features with ICC < 0.85 were excluded; (3) The minimum redundancy-maximum relevance (mRMR) algorithm was applied to remove redundant and irrelevant features; and (4) The least absolute shrinkage and selection operator (LASSO) regression was used to identify the optimal feature subset within the training dataset. The selected features were then linearly combined and weighted by their respective LASSO coefficients to calculate a radiomics score (rad-score) for each patient, which was used to construct the radiomics model (R).

DL radiomics model construction

To address the limitations of a small sample size, DL features were extracted using a 3D ResNet-18 convolutional neural network model pre-trained on ImageNet. Based on the training cohort data, all DL features were standardized through z-score normalization. Feature selection involved ICC analysis, mRMR, and LASSO regression to identify the most relevant features. The DL-score was then calculated by linearly combining these selected features weighted by their LASSO coefficients. This process resulted in the construction of a DL radiomics (DLR) model.

Construction of clinical and combined models

First, univariate analysis was conducted on all clinical risk factors. Factors with a significance level of P < 0.100 were then included in a multivariate analysis to identify independent risk factors (P < 0.05). A clinical model was subsequently developed based on these independent predictors. By integrating radiomics and DL features, a DL-radiomics integrated model (DLRR) was constructed using logistic regression. Finally, a comprehensive clinical-DL-radiomics combined model was established by combining all three components.

Model performance evaluation

Receiver operating characteristic curves and the corresponding area under the curve (AUC) values were used to assess the discriminative performance of the developed models across the training, internal validation, and external validation cohorts. Accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated based on the cutoff value determined by the Youden index in the training cohort. Following a comprehensive comparison of model performances, a nomogram was constructed using the optimal predictive model to facilitate clinical application. Calibration of the nomogram was evaluated using calibration curves and the Hosmer-Lemeshow goodness-of-fit test. Additionally, decision curve analysis was performed to assess the clinical utility of the nomogram.

Follow-up

Patients underwent regular follow-up after surgery. Those with postoperative pathology of T1-2N0M0 did not receive adjuvant therapy. Patients classified as T3-4N0M0 or T1-4N1-2M0 were administered adjuvant chemotherapy based on the XELOX regimen (capecitabine plus oxaliplatin). Recurrence was evaluated every 3 months to 6 months postoperatively through clinical assessments and imaging studies. The primary endpoint of the study was 3-year RFS. Follow-up concluded on December 31, 2024.

Statistical analysis

Statistical analyses were performed using IBM SPSS Statistics (version 29.0; IBM Corp., New York, NY, United States) and R software (version 4.3.2; http://www.r-project.org). Continuous variables are expressed as either means ± SD or medians with interquartile ranges, while categorical variables are shown as n (%). Independent sample t-tests were used for normally distributed variables, while the Mann-Whitney U test was applied to variables with skewed distributions. Categorical variables were analyzed using the χ² test or Fisher’s exact test, as appropriate. A two-tailed P value < 0.05 was considered statistically significant.

RESULTS

Patient’s characteristics at baseline

A total of 278 CRC patients who underwent surgical treatment and were pathologically confirmed postoperatively were included in this study. The clinical characteristics of all patients are summarized in Table 2. Among them, 107 patients developed LNM. There were no statistically significant differences among the training, internal validation, and external validation cohorts in terms of sex, age, height, weight, body mass index, location, low-density lipoprotein, C-reactive protein, carcinoembryonic antigen (CEA), and carbohydrate antigen 19-9 (CA 19-9). In the training cohort, CEA and CA 19-9 levels differed significantly between patients with and without LNM, suggesting that elevated CEA and CA 19-9 levels may be associated with LNM. However, other factors, including age, sex, and body mass index, showed no significant correlation with LNM in any cohort.

Table 2 Baseline characteristics of the study population, n (%).

Characteristic	Training cohort (n = 154)			Internal validation cohort (n = 66)			External validation cohort (n = 58)			P value
	LNM (n = 54)	non-LNM (n = 100)	P value	LNM (n = 29)	non-LNM (n = 37)	P value	LNM (n = 24)	non-LNM (n = 34)	P value
Sex			0.493	-		0.530	-		0.286	0.112
Male	37 (68.5)	63 (63.0)		15 (51.7)	22 (59.5)		14 (58.3)	15 (44.1)
Female	17 (31.5)	37 (37.0)		14 (48.3)	15 (40.5)		10 (41.7)	19 (55.9)
Age (years), median (IQR)	64.50 (57.00, 74.00)	69.50 (61.00,77.00)	0.054	70.00 (58.50, 72.50)	72.00 (62.50, 78.50)	0.047	66.50 (58.25, 71.00)	67.00 (62.75, 72.00)	0.448	0.203
Height (cm), median (IQR)	165.00 (160.00, 170.25)	165.00 (160.00, 170.00)	0.968	165.00 (159.00,171.00)	165.00 (158.00, 172.00)	0.801	161.50 (156.75, 170.00)	164.00 (155.00, 168.00)	0.776	0.245
Weight (kg), median (IQR)	62.00 (54.75, 68.25)	62.00 (55.00, 68.00)	0.847	65.00 (61.00, 70.00)	60.00 (53.00, 67.50)	0.108	60.50 (52.50, 69.50)	60.50 (54.50, 66.50)	0.937	0.424
BMI, median (IQR)	22.98 (20.42, 24.37)	22.63 (20.55, 24.65)	0.913	23.44 (22.22, 24.67)	22.21 (20.13, 25.37)	0.272	23.14 (20.70, 24.93)	22.99 (20.68, 24.93)	0.819	0.616
Location			0.112	-		0.375	-		0.314	0.254
RCC	21 (38.9)	23 (23.0)		6 (20.7)	8 (21.6)		6 (25.0)	6 (17.6)
LCC	14 (25.9)	31 (31.0)		9 (31.0)	17 (45.9)		8 (33.3)	7 (20.6)
RC	19 (35.2)	46 (46.0)		14 (48.3)	12 (3.24)		10 (41.7)	21 (61.8)
CRP (mg/L), median (IQR)	2.06 (0.53, 7.02)	3.19 (0.50, 11.82)	0.740	2.39 (0.63, 8.17)	5.00 (0.67, 8.43)	0.725	5.00 (0.66, 10.99)	0.87 (0.50, 8.39)	0.068	0.595
LDL (mmol/L), median (IQR)	2.13 (1.61, 2.85)	2.21 (1.85, 2.90)	0.153	2.40 (1.91, 2.98)	1.97 (1.55, 2.41)	0.041	2.07 (1.72, 2.70)	2.85 (2.24, 3.62)	0.011	0.058
CEA (ng/mL)			< 0.001	-		0.366	-		0.467	0.678
≤ 5	27 (50.0)	77 (77.0)		19 (65.5)	28 (75.7)		14 (58.3)	23 (67.6)
> 5	27 (50.0)	23 (23.0)		10 (34.5)	9 (24.3)		10 (41.7)	11 (32.4)
CA 19-9 (U/mL)			< 0.001	-		0.035	-		0.651	0.333
≤ 37	34 (63.0)	91 (91.0)		21 (72.4)	34 (91.9)		21 (87.5)	31 (91.2)
> 37	20 (37.0)	9 (9.0)		8 (27.6)	3 (8.1)		3 (12.5)	3 (8.8)

LNM: Lymph node metastasis; IQR: Interquartile range; BMI: Body mass index; RCC: Right-sided colon cancer; LCC: Left-sided colon cancer; RC: Rectal cancer; CRP: C-reactive protein; LDL: Low-density lipoprotein; CEA: Carcinoembryonic antigen; CA 19-9: Carbohydrate antigen 19-9.

Construction of radiomics, DLR, and combined models

For feature extraction, a total of 12388 intratumoral and peritumoral-3 mm radiomics features were extracted from each 3D segmentation using Pyradiomics. Following a three-step feature selection process, comprising ICC, mRMR, and LASSO regression, 7 intratumoral and 7 peritumoral radiomics features were retained (Figure 3). Additionally, 3066 intratumoral and peritumoral-3 mm DL features were extracted using ResNet-18. After applying the same three-step selection process, 12 intratumoral and 7 peritumoral-3 mm DL features were selected. Based on these features, three models were constructed: A multi-regional radiomics model, a multi-regional DLR model, and a multi-regional DLRR model. The performance of each model is presented in Figure 4. The results demonstrated that the DLRR model, which integrates both intratumoral and peritumoral-3 mm features, achieved the best performance in the training and internal validation cohorts, with AUCs of 0.935 [95% confidence interval (CI): 0.889-0.968] and 0.877 (95%CI: 0.789-0.951), respectively.

Figure 3 Radiomics and deep learning feature selection. Radiomics feature selection using least absolute shrinkage and selection operator logistic regression: A: Selection of the tuning parameter; B: Least absolute shrinkage and selection operator coefficient profiles of the radiomics features.

Figure 4 Comparative analysis of area under the receiver operating characteristic curve for the nine models. A: Comparison of the area under the receiver operating characteristic curve of the nine models in the training cohort; B: Comparison of the area under the receiver operating characteristic curve of the nine models in the internal validation cohort. ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve; T: The intratumoral model; P3: The peritumoral-3mm model; TP3: The intra-peritumoral-3mm model; R: The radiomics model; DLR: The deep learning radiomics model; DLRR: Deep learning-radiomics integrated model.

Construction of clinical model and combined model

Univariate and multivariate regression analyses identified CEA and CA 19-9 as independent risk factors for LNM. Consequently, a clinical model was constructed based on these two factors (Table 3). Subsequently, the clinical factors were combined with radiomics and DL features to develop a combined predictive model. The performance of the clinical, DLRR, and combined models is presented in Figure 5 and Table 4. The combined model demonstrated superior predictive performance, with AUC values of 0.944 (95%CI: 0.905-0.973), 0.878 (95%CI: 0.788-0.949), and 0.855 (95%CI: 0.737-0.946) in the training, internal validation, and external validation cohorts, respectively. According to the DeLong test, the differences in AUCs between the combined model and the clinical model, as well as between the DLRR and clinical models, were statistically significant (P < 0.01 for all cohorts). However, no significant difference was observed between the combined model and the DLRR model in the training, internal validation, and external validation cohorts (P = 0.238, 0.940, and 0.630, respectively). Detailed classification results for each model are visualized using confusion matrices (Figure 6).

Figure 5 Receiver operating characteristic curves of the three models. A-C: Receiver operating characteristic curves of the clinical, deep learning-radiomics integrated model and combined model for predicting lymph node metastasis in the training cohort (A), internal validation cohort (B), and external validation cohort (C), respectively. ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve; DLRR: Deep learning-radiomics integrated model; Combined: The clinical-deep learning-radiomics model; CI: Confidence interval.

Figure 6 Confusion matrix of patient-level models in the training, internal validation and external validation cohorts, respectively. A-C: Clinical model; D-F: Deep learning-radiomics integrated model; G-I: Combined model. Positive represents lymph node metastasis status; negative represents non-lymph node metastasis status. DLRR: Deep learning-radiomics integrated model; Combined: The clinical-deep learning-radiomics model.

Table 3 Univariate and multivariate logistic regression analysis.

Variable	Univariate analysis		Multivariate analysis
	OR (95%CI)	P value	OR (95%CI)	P value
Sex
Male	1 (Ref.)	-	-	-
Female	0.782 (0.387-1.581)	0.494	-	-
Age	0.975 (0.949-1.002)	0.070	0.982 (0.953-1.012)	0.235
Height	1.009 (0.962-1.060)	0.703	-	-
Weight	0.997 (0.961-1.034)	0.865	-	-
BMI	0.972 (0.869-1.008)	0.624	-	-
Location
RCC	1 (Ref.)	-	1 (Ref)	-
LCC	0.495 (0.208-1.175)	0.111	0.552 (0.227-1.342)	0.190
RC	0.452 (0.204-1.004)	0.051	0.583 (0.228-1.492)	0.261
CRP	0.997 (0.988-1.006)	0.520	-	-
LDL	0.697 (0.445-1.094)	0.117	-	-
CEA
≤ 5	1 (Ref.)	-	1 (Ref)	-
> 5	3.348 (1.649-6.797)	< 0.001	2.597 (1.199-5.623)	0.015
CA 19-9
≤ 37	1 (Ref.)	-	1 (Ref)	-
> 37	5.948 (2.467-14.337)	< 0.001	3.704 (1.435-9.563)	0.007

OR: Odds ratios; CI: Confidence interval; BMI: Body mass index; RCC: Right-sided colon cancer; LCC: Left-sided colon cancer; RC: Rectal cancer; CRP: C-reactive protein; LDL: Low-density lipoprotein; CEA: Carcinoembryonic antigen; CA 19-9: Carbohydrate antigen 19-9.

Table 4 Comparison of the performance of three models in the training, internal validation and external validation cohorts.

Cohort	Models	AUC	ACC	SEN	SPE	PPV	NPV
Training cohort	Clinical	0.703 (0.623-0.780)	0.695	0.630	0.730	0.557	0.785
	DLRR	0.935 (0.889-0.968)	0.870	0.852	0.880	0.793	0.917
	Combined	0.944 (0.905-0.973)	0.812	0.926	0.750	0.667	0.949
Internal validation cohort	Clinical	0.637 (0.527-0.751)	0.636	0.517	0.730	0.600	0.659
	DLRR	0.877 (0.789-0.951)	0.742	0.621	0.838	0.750	0.738
	Combined	0.878 (0.788-0.949)	0.773	0.793	0.757	0.719	0.824
External validation cohort	Clinical	0.624 (0.492-0.760)	0.638	0.583	0.676	0.560	0.697
	DLRR	0.849 (0.731-0.941)	0.759	0.875	0.676	0.656	0.885
	Combined	0.855 (0.737-0.946)	0.690	0.917	0.529	0.579	0.900

AUC: Area under the receiver operating characteristic curve; ACC: Accuracy; SEN: Sensitivity; SPE: Specificity; PPV: Positive predictive value; NPV: Negative predictive value; DLRR: Deep learning-radiomics integrated model.

Construction and validation of personalized nomogram

To visualize the combined model, a DLRN was constructed. Calibration curves were plotted to assess the agreement between predicted and observed outcomes (Figure 7A). The Hosmer-Lemeshow goodness-of-fit test demonstrated that the DLRN exhibited good calibration in the training, internal validation, and external validation cohorts, with P values of 0.775, 0.242, and 0.116, respectively (Figure 7B-D). Decision curve analysis further indicated that, across relevant threshold probabilities, the DLRN provided a greater net clinical benefit compared to other models (Figure 7E-G).

Figure 7 The clinical-deep learning-radiomics nomogram construction and performance evaluation. A: Is a nomogram for individual prediction of lymph node metastasis risk combined with the deep learning-radiomics integrated model and independent clinical features; B-D: Are the calibration curves of the clinical-deep learning-radiomics nomogram in the training, internal validation and external validation cohorts, respectively; E-G: Are the decision curves of the clinical-deep learning-radiomics nomogram in the training, internal validation and external validation cohorts, respectively. CA 19-9: Carbohydrate antigen 19-9; CEA: Carcinoembryonic antigen; DLRR: Deep learning-radiomics integrated model.

Interpretability analysis of clinical-DL-radiomics combined model

To evaluate the contribution of each feature to the model, SHapley Additive exPlanations (SHAP) were used to quantify the impact of individual features on the model’s predictions. In the overall visualization, the SHAP beeswarm plot (Figure 8A) highlighted the top 20 features contributing most to the model, with the four strongest predictors being CA 19-9, D_3mm_feature_59, CEA, and A_3mm_logarithm_glcm_Imc2. The SHAP heatmap (Figure 8B) illustrated the direction and magnitude of each feature’s influence across all instances in the dataset. For individual case interpretation, waterfall and force plots (Figure 9) demonstrated two representative examples of accurate predictions for LNM and non-LNM cases. Figure 9A and C depict a patient confirmed as non-LNM by postoperative pathology. The model assigned this patient a 74.4% probability of non-LNM, significantly higher than the baseline probability of 19.7%. Conversely, Figure 9B and D show a patient with pathologically confirmed LNM, for whom the model predicted a 62.6% probability of LNM, also substantially exceeding the baseline probability. The waterfall plots (Figure 9A) detail the individual contribution of each feature to the final prediction, while the force plots (Figure 9C and D) visually represent the direction and magnitude of each feature’s influence on the model’s output for these cases.

Figure 8 Global model explanation by the SHapley Additive exPlanation method. A: The SHapley Additive exPlanation beeswarm plot shows the positive or negative effects of each feature on the prediction probability through red and blue colors; B: The SHapley Additive exPlanation heatmap plot shows the direction and intensity of influence for each feature of all cases in the model. SHAP: SHapley Additive exPlanation; CA 19-9: Carbohydrate antigen 19-9; CEA: Carcinoembryonic antigen.

Figure 9 Local model explanation by the SHapley Additive exPlanation method. A and B: Shows examples of correctly predicted lymph node metastasis-negative cases; C and D: Shows examples of correctly predicted lymph node metastasis-positive cases. CA 19-9: Carbohydrate antigen 19-9; CEA: Carcinoembryonic antigen.

Survival analysis

Kaplan-Meier survival curve analysis was conducted using the receiver operating characteristic cutoff value corresponding to the maximum Youden index (0.406) as the optimal threshold to stratify patients into low-risk and high-risk groups for 3-year RFS. The survival curves demonstrated that patients with low DLRN scores (≤ -0.380) had significantly better 3-year RFS compared to those with high DLRN scores (> -0.380) (Figure 10). In the training cohort, relevant factors were initially identified through univariate Cox analysis, and those exhibiting a P value < 0.100 were subsequently included in the multivariate Cox regression model, which showed that the DLRN score (hazard ratio = 2.581, 95%CI: 1.061-6.278, P = 0.037) was an independent risk factor for 3-year RFS (Figure 11). Subsequently, a prognostic model for 3-year RFS was developed based on the DLRN score, demonstrating robust performance with AUCs of 0.826, 0.788, and 0.755 in the training, internal validation, and external validation cohorts, respectively (Figure 12).

Figure 10 Kaplan-Meier curves of 3-year recurrence-free survival. A: Training cohort; B: Internal validation cohort; C: External validation cohort, respectively.

Figure 11 Forest plot of univariate Cox and multivariate Cox analysis based on the training cohort. HR: Hazard ratio; CI: Confidence interval; RCC: Right-sided colon cancer; LCC: Left-sided colon cancer; RC: Rectal cancer; DLRR: Deep learning-radiomics integrated model; DLRN: The clinical-deep learning-radiomics nomogram; CEA: Carcinoembryonic antigen; CA 19-9: Carbohydrate antigen 19-9.

Figure 12 The receiver operating characteristic curves of a prognostic model. A: Performance of a prognostic model in the training cohort; B: Performance of a prognostic model in the internal validation cohort; C: Performance of a prognostic model in the external validation cohort. ROC: Receiver operating characteristic. AUC: Area under the receiver operating characteristic curve; CI: Confidence interval.

DISCUSSION

In this study, we developed and validated an interpretable DLRN based on CECT for preoperative, individualized prediction of LNM in CRC patients. Using SHAP analysis, we intuitively clarified the contribution of each variable to the model’s predictions. Our results demonstrated that the combined model, which integrates CEA, CA 19-9, intratumoral and peritumoral-3 mm radiomics features, and DL features, achieved excellent performance in predicting LNM in CRC. Furthermore, the preoperative prognostic model constructed using the DLRN score also showed strong predictive ability for 3-year RFS, underscoring its potential utility in preoperative risk stratification and personalized treatment planning.

Several clinical factors have been shown to be useful for preoperative prediction of LNM in CRC patients[19,20]. Li et al[21] reported that age, sex, preoperative CA 19-9 level, and preoperative CEA level were independent risk factors for LNM. Consistent with these findings, our multivariate analysis revealed that elevated preoperative CEA levels (odds ratio = 2.597, 95%CI: 1.199-5.623, P = 0.015) and CA 19-9 levels (odds ratio = 3.704, 95%CI: 1.435-9.563, P = 0.007) were significantly associated with LNM. However, other clinical characteristics such as age and sex did not significantly contribute to LNM prediction, which may be related to the characteristics of the study population itself. Nonetheless, relying solely on clinical factors for LNM prediction remains suboptimal, with AUC values of only 0.703, 0.637, and 0.624 in the training, internal validation, and external validation cohorts, respectively. This highlights the need for more accurate predictive approaches. Radiomics offers a non-invasive means to identify tumor biological behavior by extracting and analyzing high-throughput quantitative data from medical images. Recently, an increasing number of radiomics models have been developed to predict LNM in CRC patients[22]. For example, Niu et al[23] constructed a CECT-based radiomics model for LNM prediction in rectal cancer, achieving AUCs of 0.886 and 0.711 in the training and validation cohorts, respectively, significantly outperforming their clinical model, which had AUCs of 0.699 and 0.657. Although their study lacked external validation, it corroborates the value of radiomics in LNM prediction, consistent with the findings of our study.

Unlike the low-level image features extracted by traditional radiomics, DL, as an advanced machine learning technique, can automatically extract high-level, complex features from images in a data-driven manner, thereby significantly enhancing model predictive performance[24]. In this study, we employed a DL approach based on the ResNet-18 architecture for feature extraction. Compared to other classic DL architectures such as Visual Geometry Group network and Inception, ResNet-18 offers higher computational efficiency and reduces the risk of overfitting, particularly with smaller sample sizes, thus greatly decreasing the time and cost associated with model training[25,26]. The DLR model we developed, based on DL features, generally achieved higher AUCs in both the training and internal validation cohorts compared to the radiomics-only model, indicating superior performance, consistent with previous studies[27,28]. However, in the training cohort for the peritumoral-3 mm region, the DLR model’s AUC (0.773) was slightly lower than that of the radiomics model (AUC 0.786). This may be attributable to the introduction of noise when expanding the tumor boundary, especially near the tumor margins where pixel values can be affected by image resolution and acquisition parameters, which in turn may have impaired the DLR model’s performance. Subsequently, we combined radiomics features with DL features, which further enhanced the model’s predictive capability and confirmed the complementary strengths of both approaches in improving robustness and generalizability.

Recent studies have demonstrated that multi-regional radiomics can more comprehensively capture tumor heterogeneity[29-31]. In this study, we selected a 3 mm peritumoral region surrounding the tumor based on prior research[32]. Our results showed that the combined model integrating both intratumoral and peritumoral radiomics features significantly outperformed models using either intratumoral or peritumoral features alone. This finding aligns with the study by Liang et al[33]. Notably, the peritumoral-3 mm model we constructed showed substantially better performance than the intratumoral model, consistent with Wang et al[34], suggesting that features extracted from the peritumoral region may be more informative for preoperative LNM prediction than intratumoral features. This was further supported by SHAP analysis, where 7 of the top 10 most important features in the SHAP beeswarm plot were derived from the peritumoral-3 mm region, compared to only 1 from the intratumoral region. We hypothesize that this is because the peritumoral region contains diverse cellular and vascular components, which contribute to a tumor microenvironment that promotes invasion and LNM through processes such as epithelial-mesenchymal transition, angiogenesis, immune evasion, and extracellular matrix remodeling.

Previous research has demonstrated that integrating multiple relevant factors can lead to more robust predictive models[35,36]. Accordingly, we developed a multi-feature combined model by integrating clinical factors, multi-regional radiomics features, and multi-regional DL features. This combined model achieved AUCs of 0.944, 0.878, and 0.855 in the training, internal validation, and external validation cohorts, respectively, significantly outperforming both the clinical model (AUCs of 0.703, 0.637, and 0.624) and the DLRR (AUCs of 0.935, 0.877, and 0.849). These results indicate that combining features from diverse data sources enables a more comprehensive characterization of tumor biology, thereby enhancing the accuracy and robustness of LNM prediction. This approach offers stronger support for clinical individualized treatment decision-making.

Machine learning techniques are often regarded as “black boxes” due to their lack of interpretability. To address this, we employed the SHAP method combined with logistic regression to quantify and visualize the contribution of each feature to the model’s predictions. This enhances model transparency and allows clinicians to intuitively understand the relative importance of each variable in predicting LNM, facilitating focused clinical attention on key predictors. Furthermore, individual case analyses illustrated the specific impact of each variable in individual cases, enabling clinicians to more precisely assess LNM risk and tailor individualized treatment strategies.

Another noteworthy finding is that the DLRN not only demonstrated excellent performance in preoperative LNM prediction but was also significantly associated with 3-year RFS. Kaplan-Meier analysis revealed that patients with low DLRN scores had significantly better 3-year RFS compared to those with high scores. Multivariate Cox regression confirmed the DLRN score as an independent prognostic factor for 3-year RFS (hazard ratio = 1.748, 95%CI: 1.137-2.689, P = 0.011), with AUCs of 0.826, 0.788, and 0.755 in the training, internal validation, and external validation cohorts, respectively. Thus, the DLRN model not only accurately predicts preoperative LNM status but also reflects patients’ long-term prognosis through a quantified risk score, providing valuable guidance for preoperative risk stratification and personalized treatment planning.

This study has several limitations. First, its retrospective design may introduce selection bias, necessitating validation through prospective studies. Second, data were collected from only two medical centers, and the external validation cohort was relatively small. Therefore, in future research, we will incorporate data from additional centers to reduce the impact of insufficient sample size on model accuracy and enhance the model’s generalizability. Third, we included only common and easily accessible clinical variables; future studies should expand the variable range and include more clinical factors to further improve predictive performance and clinical applicability.

CONCLUSION

In conclusion, this study demonstrated that a nomogram integrating clinical features, intratumoral and peritumoral-3 mm radiomics features, and DL features effectively predicts LNM in CRC patients. The use of SHAP analysis enhanced the model’s interpretability and clinical utility. Furthermore, the prognostic model based on the DLRN score showed strong performance in predicting 3-year RFS. These findings hold promise for preoperative identification of high-risk patients and support personalized treatment decision-making.

ACKNOWLEDGEMENTS

Dr. Duan Bo, Dr. Zhang Zhenhua, and Dr. Wang Linlin (Department of Radiology, The First People’s Hospital of Hefei) are acknowledged for their expert guidance in tumor region of interest delineation. We gratefully recognize the patients who participated in this study.