Abstract
Background
Colon cancer, a frequently encountered malignancy, exhibits a comparatively poor survival prognosis. Perineural invasion (PNI), highly correlated with tumor progression and metastasis, is a substantial effective predictor of stage II-III colon cancer. Nonetheless, the lack of effective and facile predictive methodologies for detecting PNI prior operation in colon cancer remains a persistent challenge.
Method
Pre-operative computer tomography (CT) images and clinical data of patients diagnosed with stage II-III colon cancer between January 2015 and December 2023 were obtained from two sub-districts of Sun Yat-sen Memorial Hospital (SYSUMH). The LASSO/RF/PCA filters were used to screen radiomics features and LR/SVM models were utilized to construct radiomics model. A comprehensive model, shown as nomogram finally, combining with radiomics score and significant clinical features were developed and validated by area under the curve (AUC) and decision curve analysis (DCA).
Result
The total cohort, comprising 426 individuals, was randomly divided into a development cohort and a validation cohort as a 7:3 ratio. Radiomics scores were extracted from LASSO-SVM models with AUC of 0.898/0.726 in the development and validation cohorts, respectively. Significant clinical features (CA199, CA125, T-stage, and N-stage) were used to establish combining model with radiomics scores. The combined model exhibited superior reliability compared to single radiomics model in AUC value (0.792 vs. 0.726, p = 0.003) in validation cohorts. The radiomics-clinical model demonstrated an AUC of 0.918/0.792, a sensitivity of 0.907/0.813 and a specificity of 0.804/0.716 in the development and validation cohorts, respectively.
Conclusion
The study developed and validated a predictive nomogram model combining radiomics scores and clinical features, and showed good performance in predicting PNI pre-operation in stage II-III colon cancer patients.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12885-024-12951-x.
Keywords: Radiomics, PNI, Colon cancer, Predictive model
Introduction
Colorectal cancer, among the most prevalent malignancies affecting the digestive system, registered an estimated global incidence of 1.9 million diagnoses and over 904,200 fatalities in 2022, which accounts for the third incidence and the second mortality in cancers [1, 2]. Perineural invasion (PNI), a manifestation of the interaction between cancer cells and neurons, is recognized as the fifth mechanism of cancer metastasis [3–6], which may be a partial or potential cause of distant metastasis without lymphatic or vascular invasion [4, 5, 7, 8]. As a significant pathological feature of various malignancies, such as pancreatic cancer and cutaneous squamous cell carcinoma, PNI is closely associated with poor prognosis in colon cancer as well [3, 5–7, 9–13]. Patients with stage II-III colon cancer who have PNI are considered to face a higher risk of recurrence and are therefore recommended to receive adjuvant chemotherapy (ACT). In contrast, patients with stage II colon cancer without PNI do not require ACT [1, 14, 15]. Additionally, some studies in stage III colon cancer with PNI suggest that ACT may help mitigate the adverse effects of PNI and improve the survival rate, further strengthening the necessity of implementing ACT in patients with stage III colon cancer who have PNI [14, 15]. Therefore, the occurrence of PNI has a significant impact on clinical decision-making for colon cancer patients. However, studies about PNI prediction of stage II-III colon cancer in a preoperative and non-invasive methods are still limited.
Traditional radiology like Computed Tomography (CT) relies more on morphological features and experience of radiologists, leading to the subjectivity and lower sensitivity of examination results, which limits the application of traditional CT to detect PNI. Radiomics, which transformed image information invisible into high-throughput quantitative features, has been approved to effectively improve diagnosis accuracy, evaluated therapy response and predict prognosis in many malignant tumors [16–18]. In colon cancer, radiomics also plays an important role to predict lymph node metastasis, chemotherapy effect, prognosis and so on [19–22]. Notably, some previous studies reported radiomics model could predict PNI in various other malignancies, including rectal cancer, pancreatic cancer, cholangiocarcinoma, and others [23–27]. However, the current PNI prediction model for colon cancer is not subdivided for stage II-III patients and is mixed with rectal cancer patients, which is not an effective guide to stage II-III colon cancer patients who may be benefited for ACT [23, 26].
Hence, in this study, we are aimed to construct and verify a novel predictive radiomics model based on CT to estimate the probability of PNI pre-operation in patients with stage II-III colon cancer.
Methods
Patients and data management
The retrospective study was approved by the Ethics Committee of Sun Yat-sen Memorial Hospital (SYSKY-2024-278-01), and exempted the requirement for informed consent. This investigation encompassed 572 patients who underwent histopathological diagnosis of stage II-III colon cancer following radical surgery between January 2015 and December 2023 at two districts of SYSUMH (North district and South district), which has their own medical team, testing equipment and patient source. Additionally, these patients underwent abdominal contrast CT two weeks prior to surgery. The exclusion criteria were rigorously applied, comprising the following four points: (1) patients diagnosed with stage I or IV colon cancer; (2) patients with insufficient image quality in CT scans; (3) patients with other malignant tumor; (4) patients who underwent neoadjuvant chemotherapy prior to radiological assessment. Finally, totally 426 patients, 247 of which were from the north district while 179 were from the south district, were enrolled into final analysis, and the whole cohort was randomly divided into the development cohort (n = 299) and validation cohort (n = 127) as 7:3 ratio (Fig. 1).
Fig. 1.
Workflow of the study
Several clinical characteristics were collected before surgery, such as age, gender and tumor markers. The TNM staging and tumor location were documented based on the CT assessment results pre-operation. TNM staging and the diagnosis of colon cancer were conducted in strict accordance with the colorectal cancer guidelines issued by the American Joint Committee on Cancer (AJCC) and the National Comprehensive Cancer Network (NCCN) [1, 28]. The primary observational indicator of the study, PNI, was defined as either the phenomenon of colon cancer enclosing more than 33% of the circumference of a nerve or directly invading the nerve sheath [3, 5–8, 11, 12, 29–32]. All the PNI detection work was finished by two experienced pathologists separately.
Acquisition of region of interest of CT image
Contrast-enhanced abdominal and pelvic CT image acquisitions were performed preoperatively using a two 64-detector row spiral CT system (Discovery 750 HD or Revolution EVO, GE Medical System, WI, USA). CT image data were acquired 20 s and 50 s after intravenous injection of 100 ml iopromide contrast medium (Ultravist-370, Bayer Vital GmbH, Germany) at a constant 4.0 ml/s injection rate. The CT image acquisition parameters were as follows 120 kV, 200 effective mAs, 0.5 s of the gantry rotation time, and a matrix of 512 × 512. Arterial and venous phase images that reconstructed with 1.25 mm slice thickness were retrieved for quantitative analysis. All the CT data must contain the arterial phase (AP) images and venous phase (VP) images, collecting from a picture archiving and communication system (PACS) of SYSUMH.
The preoperative AP images and VP images were converted from the original DICOM format to NIFTI format using MRIcroGL software (Version 1.2.20220720) [23, 26]. The NIFTI images of both phases were then imported into ITK-SNAP software (Version 4.0.1) to delineate the regions of interest (ROI) [33]. All ROI were performed by two radiologists with more than 8 years of radiology experience, separately. The intraclass correlation coefficient (ICC) was used to evaluate the consistency of ROI drawings by these two radiologists. All images were normalized to a 100 bin width scale and resampled to a voxel size of 3*3*3 mm by sitkBSpline interpolation method. As a result, 1302 radiomics characteristics from each CT phase were extracted via ‘Pyradiomics’ package based on Python platform (Version 3.10.1) [34]. The radiomics features were classified into following four groups: (a) First-order statistics (n = 18); (b) shape features (n = 14); (c) Texture features (n = 74), including 23 gray level co-occurrence matrix (GLCM) features, 16 gray level run length matrix (GLRLM) features, 16 gray level size zone matrix (GLSZM) features, 5 neighboring gray tone difference matrix (NGTDM) features and 14 gray level dependence matrix (GLDM) features; (d) higher-order statistical features (n = 1196) consisting of fist-order statistics, shape features and texture features derived from wavelet filter and Laplacian of Gaussian filter (σ-1, 2, 3, 4, 5). The details of research process were shown in Fig. 2.
Fig. 2.
The process of model based on radiomics construction
Development of radiomics model
Prior to feature selection, all radiomics features were normalized by Z-score method. Subsequently, the Least Absolute Shrinkage and Selection Operator (LASSO), the Random Forest (RF) algorithm, and Principal Component Analysis (PCA) were employed to preliminarily filter the most impactful ones from a pool of 2604 CT radiomics features. Then the Logistic Regression (LR) and Support Vector Machine (SVM) were used to construct prediction models, respectively. These models were designated as LASSO-LR, LASSO-SVM, PCA-LR, PCA-SVM, RF-LR, and RF-SVM. To assess the accuracy and efficacy of these models, a comparative analysis was conducted, leading to the selection of the best model. Finally, utilizing this optimal model, a radiomics score was computed for further analysis and application.
Establishment of the prediction model
After rigorous statistical analysis, univariate and multivariate logistic regression were performed on clinical data and radiomics score. Significant risk factors with p-values less than 0.05 were identified and subjected to further model establishment by Logistics regression. The area under the curve (AUC) of receiver operating characteristic (ROC) curve analysis was used to evaluate the prediction value of the combination model and radiomics model, and the decision curve analysis (DCA) was applied to estimate the benefits to clinical utility of prediction model. The final model to predict PNI was shown as nomogram.
Statistics analyses
In this study, all model construction, comparison, and risk factor screening were accomplished utilizing the R software (version 4.3) and RStudio (version 2023.6.0). Statistical analysis was conducted with the assistance of SPSS 25.0 software. Continuous variables were presented as mean ± standard deviation, while categorical variables were exhibited as frequency (percentage). Depending on the grouping of the two cohorts, methods such as independent sample t-test, Mann-Whitney U test, chi-square test, and Fisher’s exact test were employed for comparing differences. The comparison of AUC was achieved through the Delong’s test. The reported p-values in this study were all based on two-sided tests, and a significant difference was indicated by p < 0.05.
Results
Clinical characteristics
Totally, 426 patients, 107 of which had PNI in pathological reports, diagnosed as stage II-III colon cancer were enrolled into this study. These 426 patients were randomly divided into two groups, namely the development group and the validation group, according to a ratio of 7:3. The mean age of the whole cohort was 60 (24–87) years old and the gender distribution within the cohort leaned towards a ratio of approximately 6:4 (male: female). The baseline characteristics of two groups as well as the comparison between characteristics of PNI negative and positive were shown in Table 1and Table 2, which supported the absence of significant differences between two cohorts (p > 0.05) and detailed distribution of variables in patients with or without PNI.
Table 1.
Clinical and pathological features of development and validation cohort
| Development cohort N = 299 |
Validation cohort N = 127 |
P value | |
|---|---|---|---|
| Age (y) | 59.8 ± 12.3 | 61.1 ± 13.9 | 0.384 |
| Gender | 0.452 | ||
| Male (%) | 175 (58.5%) | 80 (63.0%) | |
| Female (%) | 124 (41.5%) | 47 (37.0%) | |
| Location | 0.714 | ||
| Ileocecal (%) | 3 (1.00%) | 2 (1.57%) | |
| Ascending (%) | 81 (27.1%) | 27 (21.3%) | |
| Transverse (%) | 28 (9.36%) | 14 (11.0%) | |
| Descending (%) | 34 (11.4%) | 16 (12.6%) | |
| Sigmoid (%) | 153 (51.2%) | 68 (53.5%) | |
| CA199 (U/mL) | 129 ± 807 | 122 ± 580 | 0.926 |
| CEA (ng/mL) | 38.9 ± 287 | 220 ± 1991 | 0.311 |
| AFP (ng/mL) | 3.23 ± 6.97 | 2.56 ± 1.15 | 0.112 |
| CA125 (U/mL) | 20.9 ± 28.1 | 22.2 ± 36.1 | 0.724 |
| T stage | 0.662 | ||
| T2-T3 (%) | 185 (61.9%) | 75 (59.1%) | |
| T4 (%) | 114 (38.1%) | 52 (40.9%) | |
| N stage | 0.843 | ||
| N0 (%) | 113 (37.8%) | 50 (39.4%) | |
| N1-N2 (%) | 186 (62.2%) | 77 (60.6%) | |
| Differentiation | 0.838 | ||
| Low (%) | 41 (13.7%) | 16 (12.6%) | |
| Middle (%) | 243 (81.3%) | 103 (81.1%) | |
| High (%) | 15 (5.02%) | 8 (6.3%) | |
| PNI | 0.379 | ||
| (-) (%) | 228 (76.3%) | 91 (71.7%) | |
| (+) (%) | 71 (23.7%) | 36 (28.3%) |
Location, the position of colon cancer, from ileocecal colon to sigmoid colon; CA199, carbohydrate antigen 19 − 9; CA125, carbohydrate antigen 125; CEA, carcinoembryonic antigen; AFP, alpha fetoproteinl; T stage and N stage, T and N staging based on CT images according to AJCC 8th Edition; Differentiation, evaluated and reported by pathologists, including low differentiated, middle differentiated, high differentiated; PNI, perineural invasion
Table 2.
Difference of features between PNI (-) or (+) in development and validation cohort
| Characteristics | Development cohort | Validation cohort | ||||
|---|---|---|---|---|---|---|
| PNI ( - ) | PNI ( + ) | P value | PNI ( - ) | PNI ( + ) | P value | |
| N = 228 | N = 71 | N = 91 | N = 36 | |||
| Age (y) | 60.1 ± 11.8 | 59.0 ± 13.7 | 0.571 | 61.1 ± 13.3 | 61.0 ± 15.6 | 0.990 |
| Gender | 1.000 | 0.942 | ||||
| Male (%) | 133 (58.3%) | 42 (59.2%) | 58 (63.7%) | 22 (61.1%) | ||
| Female (%) | 95 (41.7%) | 29 (40.8%) | 33 (36.3%) | 14 (38.9%) | ||
| Location | 0.229 | 0.447 | ||||
| Ileocecal (%) | 2 (0.88%) | 1 (1.41%) | 1 (1.10%) | 1 (2.78%) | ||
| Ascending (%) | 62 (27.2%) | 19 (26.8%) | 16 (17.6%) | 11 (30.6%) | ||
| Transverse (%) | 17 (7.46%) | 11 (15.5%) | 11 (12.1%) | 3 (8.33%) | ||
| Descending (%) | 25 (11.0%) | 9 (12.7%) | 12 (13.2%) | 4 (11.1%) | ||
| Sigmoid (%) | 122 (53.5%) | 31 (43.7%) | 51 (56.0%) | 17 (47.2%) | ||
| CA199 (U/mL) | 53.3 ± 213 | 371 ± 1595 | 0.099 | 69.0 ± 418 | 257 ± 859 | 0.216 |
| CEA (ng/mL) | 12.8 ± 21.8 | 123 ± 582 | 0.116 | 259 ± 2345 | 119 ± 340 | 0.578 |
| AFP (ng/mL) | 3.33 ± 7.96 | 2.87 ± 1.30 | 0.400 | 2.54 ± 1.18) | 2.61 ± 1.08 | 0.762 |
| CA125 (U/mL) | 17.3 ± 22.3 | 32.5 ± 39.6 | 0.003 | 19.7 ± 32.6 | 28.4 ± 43.5 | 0.286 |
| T stage | 0.004 | 0.481 | ||||
| T2-T3 (%) | 152 (66.7%) | 33 (46.5%) | 56 (61.5%) | 19 (52.8%) | ||
| T4 (%) | 76 (33.3%) | 38 (53.5%) | 35 (38.5%) | 17 (47.2%) | ||
| N stage | <0.001 | 0.060 | ||||
| N0 (%) | 102 (44.7%) | 11 (15.5%) | 41 (45.1%) | 9 (25.0%) | ||
| N1-N2 (%) | 126 (55.3%) | 60 (84.5%) | 50 (54.9%) | 27 (75.0%) | ||
| Differentiation | 0.002 | 0.058 | ||||
| Low (%) | 24 (10.5%) | 17 (23.9%) | 9 (9.89%) | 7 (19.4%) | ||
| Middle (%) | 189 (82.9%) | 54 (76.1%) | 74 (81.3%) | 29 (80.6%) | ||
| High (%) | 15 (6.58%) | 0 (0.00%) | 8(8.79%) | 0 (0.00%) | ||
| Radio-score | 0.20 ± 0.13 | 0.41 ± 0.19 | <0.001 | 0.21 ± 0.13 | 0.46 ± 0.15 | <0.001 |
Radio-score, the likelihood of PNI occurrence in stage II-III colon cancer
Prediction of PNI with radiomics characteristics
Upon comparing ROI and CT images, 2604 radiomics features, with the ICC of 0.842 that was highly consistent, were analyzed to construct six radiomics models using various statistical methods (Table S2). Among these models, the LASSO-SVM model exhibited a sensitivity of 0.893/0.719 and specificity of 0.804/0.705, achieving an AUC of 0.898/0.726 in the development and validation cohorts, respectively (Fig. 4A). In contrast, the LASSO-LR model demonstrated an AUC of 0.901/0.725 (sensitivity: 0.893/0.343; specificity: 0.790/0.916); the RF-SVM model yielded an AUC of 0.694/0.522 (sensitivity: 0.707/0.406; specificity: 0.652/0.590); the RF-LR model obtained an AUC of 0.690/0.463 (sensitivity: 0.747/0.031; specificity: 0.549/0.978); the PCA-SVM model showed an AUC of 0.522/0.542 (sensitivity: 0.533/0.656; specificity: 0.375/0.473); and the PCA-LR model had an AUC of 0.597/0.555 (sensitivity: 0.707/0.001; specificity: 0.500/0.999). Notably, the LASSO-SVM model outperformed the other models, demonstrating superior AUC, sensitivity, and specificity in both cohorts simultaneously (Table S1). The results of LASSO regression for the LASSO-SVM model are presented in Fig. 3. Utilizing this model, a predictive score for PNI in stage II-III colon cancer, termed “radio-score”, was easily calculated (Fig. 4B-C). The radio-score represents the likelihood of PNI occurrence in stage II-III colon cancer and ranges from 0 to 1 (OR = 6007; 95% CI, [1062, 40391], p < 0.001) (Table 3).
Fig. 3.
LASSO regression analysis for selection of radiomics features. A λ called penalty factor was gotten after tenfold cross-validation. λ minimum and λ 1-SE were selected to sign the dotted vertical line in the plot. B LASSO coefficient profile for predicting PNI in stage II-III colon cancer
Fig. 4.
Data of LASSO-SVM radiomics model. A ROC of LASSO-SVM radiomics model of development cohort and validation cohort, respectively. B Radio-score in stage II-III colon cancer with or without PNI. C Radio-score in high or low risk of PNI in stage II-III colon cancer
Table 3.
Univariate and Multivariate logistic regression analysis of clinical and pathological features
| Characteristic | Univariate logistic regression | |
|---|---|---|
| OR (95% CI) | P.value | |
| Age (y) | 0.996(0.979–1.013) | 0.652 |
| Gender (male/female) | 1.003(0.639–1.563) | 0.991 |
|
Tumor position (lleocecal-sigmoid) |
0.577(0.091–4.542) | 0.558 |
| CA199 (U/ml) | 1.001(1.000-1.001) | 0.036 |
| CEA (ng/ml) | 1.00002(0.9997–1.0002) | 0.761 |
| AFP (n/ml) | 0.984(0.984 − 0.862) | 0.656 |
| CA125 (U/ml) | 1.012(1.005–1.019) | 0.001 |
| T stage (T2-3/T4) | 0.952(0.320–2.875) | 0.003 |
| N stage (N0/N1-2) | 1.982(1.272–3.095) | < 0.001 |
| Radio-score | 6007(1062–40391) | < 0.001 |
Radio-score, probability of PNI calculated with radiomics models
Clinical data selection and radiomics-clinical combining model to predict PNI
In addition to radiological data, tumor markers and T&N staging from radiology were concurrently collected. Following a univariate and multivariate logistic regression analysis of the clinical data, CA199 (p = 0.036), CA125 (p = 0.001), T-stage (p = 0.009), and N-stage (p = 0.003) were identified as significant clinical risk factors (Tables 3 and 4).
Table 4.
Risk factors for PNI in stage II-III colon cancer
| Intercept and variable | β | OR (95% CI) | P.value |
|---|---|---|---|
| Intercept | -5.261 | 0.005(0.001–0.016) | < 0.001 |
| CA199(U/ml) | 0.001 | 1.001(1.000-1.002) | 0.145 |
| CA125(U/ml) | 0.008 | 1.008(0.998–1.020) | 0.120 |
| T stage | 1.309 | 2.584(1.275–5.379) | 0.009 |
| N stage | 0.949 | 3.701(1.637–9.082) | 0.003 |
| Radio-score | 8.371 | 4318(478-51400) | < 0.001 |
β is the regression coefficient
Finally, a radiomics-clinical model, which integrates the significant clinical features with the radio-score, performed better than the radiomics model only. The combined model exhibited an AUC of 0.918/0.792, with sensitivity and specificity values of 0.907/0.813 and 0.804/0.716 in development/validation cohort, respectively. The combined model demonstrated a similar AUC with radiomics model in the development cohort (0.918 vs. 0.898, p = 0.123) but a higher AUC than radiomics model in the validation cohort (0.792 vs. 0.726, p = 0.003) (Fig. 5A-B). The DCA results revealed that the curve of the combining model usually exceeded radiomics model, indicating that decisions made based on the combining model are more reliable most of the time (Fig. 5C-D). Consequently, the radiomics-clinical model demonstrated superior performance than single radiomics models.
Fig. 5.
ROC and decision curve of radiomics-clinical and radiomics models in development and validation cohort. ROC of development cohort (A) and validation cohort (B) of 2 models. Decision curve analysis of development cohort (C) and validation cohort (D) of 2 models
The prediction model is presented in Table 4, which identifies the radio-score, radiological T stage, and radiological N stage as independent risk factors for PNI in stage II-III colon cancer. Consequently, the nomogram of the radiomics-clinical model is depicted in Fig. 6.
Fig. 6.
Nomogram of the radiomics-clinical model
Discussion
Colon cancer represents the second most prevalent cause of cancer mortality globally [35, 36]. PNI serves as an independent prognostic factor for poor outcomes in colon cancer. Its presence often prompts clinicians to recommend adjuvant chemotherapy for patients with stage II-III colon cancer [3, 4, 12, 31, 37–39]. In cases where therapeutic progress appears to stagnate, the classification of colon cancer into PNI-positive and PNI-negative subtypes carries the hope of refining our understanding of the disease’s heterogeneity and guiding more personalized treatment strategies. Notably, the radiomics-clinical models developed in this study offer a convenient and non-invasive approach to predict PNI in stage II-III colon cancer.
Our study focuses on PNI in stage II-III colon cancer while previous studies have attempted to predict PNI or other outcomes in both colon and rectal cancers, collectively referred to as colorectal cancer (CRC) [27, 40–42]. However, the distinct anatomical and clinical characteristics of colon and rectal cancers, including differences in vascular and neural distribution, suggest that the two cancers need to be studied separately. Additionally, the limitations of CT imaging in depicting radiological features of rectal cancer, due to its poor soft-tissue contrast, further support the preferential use of magnetic resonance imaging (MRI) for rectal cancer diagnosis rather than CT [10, 43, 44]. Therefore, it may not be appropriate to generalize predictions of PNI in CRC based on CT. Regarding stage IV colon cancer, the notably bad prognosis diminishes the significance of predicting PNI. Similarly, for patients with stage I colon cancer, nearly all of which survive in five years after radical surgery, ACT is generally not recommended, no matter whether PNI occurs or not. Conversely, ample evidences support the significant role of PNI in stage II-III colon cancer, influencing decisions about ACT administration and prognosis prediction [32, 38, 45–48]. Consequently, our study focused on establishing a predictive model specific to stage II-III colon cancer.
Numerous studies has proposed the prediction models of PNI with radiomics in many kinds of cancer, such as rectal cancer, cholangiocarcinoma, head and neck squamous cell carcinoma and so on [26, 49, 50]. As shown in the results, the LASSO-SVM radiomics model performed the most excellent among the six models. In details, 31 variables were filered by LASSO regression analysis from 2604 features, significantly reducing the computational complexity for SVM and improving the generalization of the model. The radiomics features were divided into semantic and agnostic features. The semantic features contain shape, size and others features describing the ROI, while the later had three orders called the first-, second- and higher-orders. All radiomics features selected in our study were the higher-order features, which were extracted from low-order features transformed by wavelet filter and Laplacian of Gaussian filter (Table S1). This indicated that the first-order features, describing the distribution of values of individual voxels without the concerns for spatial relationships, and shape features played little role in PNI prediction [18]. Some studies also support that the higher-order features accounts for more predictive value than the lower-order features [51, 52]. However, it is a dilemma that the higher order radiological data cannot be visually recognized and have no good clinical explanation, which indicates that the necessity of further research.
Besides, to improve the models, tumor markers, T stage and N stage were integrated with the radiomics model and then a combined model, demonstrating superior performance compared to the LASSO-SVM model, was developed according to the univariate logistic regression analysis. The combined model contained not only radiological information, but also tumor size, lymph node metastasis and tumor markers, providing a more comprehensive assessment of the patient’s overall condition. As a result, tumor markers seem to be associated with PNI and have a potential to enhance the PNI model based on radiomics. In a study by Huang YQ, CEA was incorporated into the model to predict PNI in colorectal cancer. However, its contribution was found to be limited [41]. Similarly, in our study, CEA exhibited a p-value of 0.761 in univariate regression analysis, with statistical significance observed only in multivariate logistic regression analysis using backward elimination. This poor performance of CEA in the model suggests its limited utility in predicting PNI, despite potential associations. When it comes to the CA125 and CA19-9 in our study, although there are few models containing these two tumor markers about colon cancer PNI prediction, their contribution for PNI prediction of colon cancer in our nomogram of combining model could not be simply overlooked. The absence of these markers in other studies may be attributed to insufficient sample sizes while we limited the enrolled patients to stage II-III colon cancer. Therefore, further researches are needed to explore the significance and roles of CA125 and CA199 in colon cancer PNI. Since the level of tumor markers may be influenced by kidney function and some other diseases beside colon cancer, the different points in terms of tumor markers among the studies may be attributed to the complex environment of patients. T stage describes the extent of colon cancer infiltration of colonic wall. The varying distribution of nerves across different layers of the colonic wall suggests a correlation between T stage and PNI. As for N stage, as an alternative mechanism for metastasis, PNI may be more likely to occur in the presence of lymph node metastasis.
Our study also encounters several limitations. Firstly, the CT scans and clinical data may lack objectivity due to the constraints of the inherent variability of CT device parameters. Secondly, the patient cohort fails to comprehensively represent the entire spectrum of age groups and disease stages in colon cancer, necessitating further investigation, particularly in the context of predicting the probability of PNI in younger and older populations. Thirdly, our study is a retrospective study, and our PNI model requires further external validation and prospective study validation.
Conclusion
We developed and validated a radiomics-clinical model, which integrates clinical data with radiomic features extracted from CT images, for predicting PNI in patients diagnosed with stage II-III colon cancer. This model facilitates the identification of PNI pre-operation, thereby aiding clinicians in tailoring more personalized treatment strategies for individuals with colon cancer.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Acknowledgements
We thank all the participants who worked for this study. We especially thank Dr. Jiang Chendan for his help for our issues of python code.
Abbreviations
- ACT
Adjuvant chemotherapy
- AFP
Alpha fetoproteinl
- AJCC
American Joint Committee on Cancer
- AUC
Area under the receiver operating characteristic curve
- CA125
Carbohydrate antigen 125
- CEA
Carcinoembryonic antigen
- CA199
Carbohydrate antigen 19 − 9
- CT
Computer tomography
- DCA
Decision curve analysis
- GLCM
Gray level co-occurrence matrix
- GLDM
Gray level dependence matrix
- GLRLM
Gray level run length matrix
- GLSZM
Gray level size zone matrix
- LASSO
Least Absolute Shrinkage and Selection Operator
- LR
Logistic Regression
- NCCN
National Comprehensive Cancer Network
- NGTDM
Neighboring gray tone difference matrix
- PACS
Picture archiving and communication system
- PCA
Principal Component Analysis
- PNI
Perineural invasion
- RF
Random Forest
- ROI
region of interest
- SVM
Support Vector Machine
- SYSUMH
Sun Yat-sen Memorial Hospital
Author contributions
GTR., CB. and LYL. were co-first authors and contributed equally to this study in study design, data acquisition, data analysis and manuscript drafting. GM., HKH. and HYZ. were the corresponding author responsible for study design, quality control and manuscript drafting. LGD. and GM. were responsible for ROIs segmentation. LYQ., CSJ. and LJJ. help to clinical data acquisition and data analysis.
Funding
This study was supported by National Natural Science Foundation of China (NSFC 82203036, 82373025), Guangzhou Science and Technology Plan Project (SL2024A04J01991), Natural Science Foundation of Guangdong Province (2024A1515012799).
Data availability
All data generated or analyzed during this study are included in this published article and its supplementary information files. The data that support the findings of this study are available from the corresponding author upon reasonable request.
Declarations
Ethics approval and consent to participate
This study was approved by the Sun Yat-sen Memorial Hospital Ethics Committee (SYSKY-2024-278-01), and informed consent from participants was exempted due to retrospective study.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Tairan Guo, Bing Cheng and Yunlong Li co-first authors and contributed equally to this study.
Contributor Information
Ming Gao, Email: gaoming2@mail.sysu.edu.cn.
Kaihong Huang, Email: huangyzh83@mail.sysu.edu.cn.
Yuzhou Huang, Email: huangkh@mail.sysu.edu.cn.
References
- 1.Benson AB, Venook AP, Bekaii-Saab T, et al. Colon cancer, Version 3.2014. J Natl Compr Canc Netw. 2014;12:1028–59. [DOI] [PubMed] [Google Scholar]
- 2.Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74:229–63. [DOI] [PubMed] [Google Scholar]
- 3.Liu Q, Ma Z, Cao Q, et al. Perineural invasion-associated biomarkers for tumor development. Biomed Pharmacother. 2022;155:113691. [DOI] [PubMed] [Google Scholar]
- 4.Wang H, Huo R, He K, et al. Perineural invasion in colorectal cancer: mechanisms of action and clinical relevance. Cell Oncol. 2024;47:1–17. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Liebig C, Ayala G, Wilks JA, et al. Perineural invasion in cancer: a review of the literature. Cancer. 2009;115:3379–91. [DOI] [PubMed] [Google Scholar]
- 6.Yang M-W, Tao L-Y, Jiang Y-S, et al. Perineural Invasion reprograms the Immune Microenvironment through Cholinergic Signaling in Pancreatic Ductal Adenocarcinoma. Cancer Res. 2020;80:1991–2003. [DOI] [PubMed] [Google Scholar]
- 7.Patel BN, Olcott E, Jeffrey RB. Extrapancreatic perineural invasion in pancreatic adenocarcinoma. Abdom Radiol. 2018;43:323–31. [DOI] [PubMed] [Google Scholar]
- 8.Schmitd LB, Scanlon CS, D’Silva NJ. Perineural Invasion in Head and Neck Cancer. J Dent Res. 2018;97:742–50. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Cohen ER, Misztal C, Dable C, et al. Redefining Perineural Invasion in Head and Neck cutaneous squamous cell carcinoma. Otolaryngol–head neck surg. 2022;167:705–15. [DOI] [PubMed] [Google Scholar]
- 10.Zhang L, Yang L, Jiang S, et al. Nerve dependence in Colorectal Cancer. Front Cell Dev Biol. 2022;10:766653. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Hu G, Li L, Hu K. Clinical implications of perineural invasion in patients with colorectal cancer. Medicine. 2020;99:e19860. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Knijn N, Mogk SC, Teerenstra S, et al. Perineural Invasion is a strong prognostic factor in Colorectal Cancer: a systematic review. Am J Surg Pathol. 2016;40:103–12. [DOI] [PubMed] [Google Scholar]
- 13.Wang H, Huo R, He K et al. Perineural invasion in colorectal cancer: mechanisms of action and clinical relevance. Cell oncol. Epub ahead of print 23 August 2023. 10.1007/s13402-023-00857-y [DOI] [PMC free article] [PubMed]
- 14.Leijssen LGJ, Deshpande V, Bordeianou LG. Perineural Invasion is a prognostic but not a predictive factor in nonmetastatic Colon cancer. Dis Colon Rectum. 2019;62(10):1212–21. [DOI] [PubMed]
- 15.Song JH, Yu M, Kang KM, et al. Significance of perineural and lymphovascular invasion in locally advanced rectal cancer treated by preoperative chemoradiotherapy and radical surgery: can perineural invasion be an indication of adjuvant chemotherapy? Radiother Oncol. 2019;133:125–31. [DOI] [PubMed] [Google Scholar]
- 16.Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48:441–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Rubin DL. Informatics in Radiology: measuring and improving quality in Radiology: meeting the challenge with Informatics. Radiographics. 2011;31:1511–27. [DOI] [PubMed] [Google Scholar]
- 18.Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures. They Are Data Radiol. 2016;278:563–77. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Defeudis A, Cefaloni L, Giannetto G et al. Comparison of radiomics approaches to predict resistance to 1st line chemotherapy in liver metastatic colorectal cancer. In: 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). Mexico: IEEE, pp. 3305–3308. [DOI] [PubMed]
- 20.Miccichè F, Rizzo G, Casà C, et al. Role of radiomics in predicting lymph node metastasis in gastric cancer: a systematic review. Front Med. 2023;10:1189740. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Abbaspour E, Karimzadhagh S, Monsef A, et al. Application of radiomics for preoperative prediction of lymph node metastasis in colorectal cancer: a systematic review and meta-analysis. Int J Surg. 2024;110:3795–813. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Yao X, Sun C, Xiong F, et al. Radiomic signature-based nomogram to predict disease-free survival in stage II and III colon cancer. Eur J Radiol. 2020;131:109205. [DOI] [PubMed] [Google Scholar]
- 23.Huang Y, Zhou S, Luo Y, et al. Development and validation of a radiomics model of magnetic resonance for predicting liver metastasis in resectable pancreatic ductal adenocarcinoma patients. Radiat Oncol. 2023;18:79. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Wang M, Sun Z, Wang Q, Guo J, Li X, Chen J, et al. Establishment and validation of a nomogram model for preoperative prediction of the risk of cholangiocarcinoma with perineural invasion. Am J Cancer Res. 2023;13(11):5082–93. [PMC free article] [PubMed]
- 25.Liu Y-L, Zhu H-B, Chen M-L, et al. Prediction of the lymphatic, microvascular, and perineural invasion of pancreatic neuroendocrine tumors using preoperative magnetic resonance imaging. World J Gastrointest Surg. 2023;15:2809–19. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Chen J, Chen Y, Zheng D, et al. Pretreatment MR-based radiomics nomogram as potential imaging biomarker for individualized assessment of perineural invasion status in rectal cancer. Abdom Radiol. 2021;46:847–57. [DOI] [PubMed] [Google Scholar]
- 27.Qian H, Huang Y, Xu L, et al. Role of peritumoral tissue analysis in predicting characteristics of hepatocellular carcinoma using ultrasound-based radiomics. Sci Rep. 2024;14:11538. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Amin MB, Greene FL, Edge SB, et al. The Eighth Edition AJCC Cancer staging Manual: continuing to build a bridge from a population-based to a more personalized approach to cancer staging. CA Cancer J Clin. 2017;67:93–9. [DOI] [PubMed] [Google Scholar]
- 29.Zhou AE, Hoegler KM, Khachemoune A. Review of Perineural Invasion in Keratinocyte Carcinomas. Am J Clin Dermatol. 2021;22:653–66. [DOI] [PubMed] [Google Scholar]
- 30.Zhang W, He R, Yang W, et al. Autophagic Schwann cells promote perineural invasion mediated by the NGF/ATG7 paracrine pathway in pancreatic cancer. J Exp Clin Cancer Res. 2022;41:48. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Cienfuegos JA, Martínez P, Baixauli J, et al. Perineural Invasion is a Major Prognostic and Predictive Factor of Response to Adjuvant Chemotherapy in Stage I–II Colon cancer. Ann Surg Oncol. 2017;24:1077–84. [DOI] [PubMed] [Google Scholar]
- 32.Chu C-H, Lai I-L, Jong B-K, et al. The prognostic and predictive significance of perineural invasion in stage I to III colon cancer: a propensity score matching-based analysis. World J Surg Onc. 2024;22:129. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Yushkevich PA, Piven J, Hazlett HC, et al. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. NeuroImage. 2006;31:1116–28. [DOI] [PubMed] [Google Scholar]
- 34.Van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic phenotype. Cancer Res. 2017;77:e104–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Bray F, Ferlay J, Soerjomataram I, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68:394–424. [DOI] [PubMed] [Google Scholar]
- 36.Li Y, Eresen A, Shangguan J, et al. Preoperative prediction of perineural invasion and KRAS mutation in colon cancer using machine learning. J Cancer Res Clin Oncol. 2020;146:3165–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Baxter NN, Kennedy EB, Bergsland E, et al. Adjuvant therapy for stage II Colon cancer: ASCO Guideline Update. JCO. 2022;40:892–910. [DOI] [PubMed] [Google Scholar]
- 38.Suzuki T, Suwa K, Ogawa M, et al. Adjuvant chemotherapy for the perineural invasion of colorectal cancer. J Surg Res. 2015;199:84–9. [DOI] [PubMed] [Google Scholar]
- 39.Mirkin KA, Hollenbeak CS, Mohamed A, et al. Impact of perineural invasion on survival in node negative colon cancer. Cancer Biol Ther. 2017;18:740–5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Ma J, Guo D, Miao W, et al. The value of 18F-FDG PET/CT-based radiomics in predicting perineural invasion and outcome in non-metastatic colorectal cancer. Abdom Radiol. 2022;47:1244–54. [DOI] [PubMed] [Google Scholar]
- 41.Huang Y, He L, Dong D, et al. Individualized prediction of perineural invasion in colorectal cancer: development and validation of a radiomics prediction model. Chin J Cancer Res. 2018;30:40–50. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Fan S, Cui X, Liu C, et al. CT-Based Radiomics signature: a potential biomarker for Predicting Postoperative Recurrence Risk in Stage II Colorectal Cancer. Front Oncol. 2021;11:644933. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43.Benson AB, Venook AP, Bekaii-Saab T, et al. Rectal Cancer, Version 2.2015. J Natl Compr Canc Netw. 2015;13:719–28. [DOI] [PubMed] [Google Scholar]
- 44.Grazzini G, Danti G, Chiti G, et al. Local recurrences in rectal Cancer: MRI vs. CT Diagnostics. 2023;13:2104. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Kang JH, Son IT, Kim BC et al. Recurrence-Free Survival Outcomes Based on Novel Classification Combining Lymphovascular Invasion, Perineural Invasion, and T4 Status in Stage II–III Colon Cancer. CMAR. 2022; Volume 14: 2031–2040. [DOI] [PMC free article] [PubMed]
- 46.Salvatore L, Imperatori M, Arnoldi E, et al. Management of patients with early-stage colon cancer: guidelines of the Italian Medical Oncology Association. ESMO Open. 2020;5:e001001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Cervantes A, Adam R, Roselló S, et al. Metastatic colorectal cancer: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up. Ann Oncol. 2023;34:10–32. [DOI] [PubMed] [Google Scholar]
- 48.Simillis C, Singh HKSI, Afxentiou T, et al. Postoperative chemotherapy improves survival in patients with resected high-risk stage II colorectal cancer: results of a systematic review and meta‐analysis. Colorectal Dis. 2020;22:1231–44. [DOI] [PubMed] [Google Scholar]
- 49.Liu Z, Luo C, Chen X, et al. Noninvasive prediction of perineural invasion in intrahepatic cholangiocarcinoma by clinicoradiological features and computed tomography radiomics based on interpretable machine learning: a multicenter cohort study. Int J Surg. 2024;110:1039–51. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 50.Mukherjee P, Cintra M, Huang C, et al. CT-based Radiomic Signatures for Predicting Histopathologic Features in Head and Neck squamous cell carcinoma. Radiology: Imaging Cancer. 2020;2:e190039. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 51.Chen Q, Cui Y, Xue T, et al. Computed tomography-based radiomics nomogram for the preoperative prediction of perineural invasion in colorectal cancer: a multicentre study. Abdom Radiol. 2022;47:3251–63. [DOI] [PubMed] [Google Scholar]
- 52.He Y, Yang M, Hou R, et al. Preoperative prediction of perineural invasion and lymphovascular invasion with CT radiomics in gastric cancer. Eur J Radiol Open. 2024;12:100550. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
All data generated or analyzed during this study are included in this published article and its supplementary information files. The data that support the findings of this study are available from the corresponding author upon reasonable request.