Nomogram including tumor deposition count to noninvasively evaluate the prognosis of rectal cancer patients: A retrospective study

Abstract

To build a nomogram model that includes tumor deposition (TDs) count to noninvasively evaluate the prognosis of patients with rectal cancer (RC). A total of 262 patients between January 2013 and December 2018 were recruited and divided into 2 cohorts: training (n = 171) and validation (n = 91). Axial portal venous phase computed tomography images were used to extract radiomic features, and the least absolute shrinkage and selection operator-Cox analysis was applied to develop an optimal radiomics model to derive the Rad-score. A Cox regression model combining clinicopathological factors and Rad-scores was constructed and visualized using a nomogram. And its ability to predict RC patients’ survival was tested by Kaplan–Meier survival analysis. The time-dependent concordance index curve was used to demonstrate the differentiation degree of model. Calibration and decision curve analyses were used to evaluate the calibration accuracy and clinical usefulness of the nomogram model, and the prediction performance of the nomogram model was compared with the clinical and radiomics models using the likelihood test. Computed tomography-based Rad-score, pathological tumor (pT) stageT4, and TDs count were independent risk factors affecting the prognosis of RC. The whole concordance index of the nomogram model for predicting the overall survival rates of RC was higher than that of the clinical and radiomics models in the training (0.812 vs 0.59, P = .019; 0.812 vs 0.714, P = .014) and validation groups (0.725 vs 0.585, P = .002; 0.725 vs 0.751, P = .256). The nomogram model could effectively predict patients’ overall survival rate (hazard ratio = 9.25, 95% CI = [1.17–72.99], P = .01). The nomogram model also showed a higher clinical net benefit than the clinical and radiomics models in the training and validation groups. The nomogram model developed in this study can be used to noninvasively evaluate the prognosis of RC patients. The TDs count is an independent risk factor for the prognosis of RC.

1. Introduction

Rectal cancer (RC) is a common gastrointestinal malignancy with an increasing incidence and mortality worldwide,^[1] causing serious threats to human health and quality of life. The prognosis of patients with RC is poor due to RC’s highly aggressive behavior nature; for example, approximately 25% of patients with early RC have distant metastasis.^[2] Early and metaphase RC lacks typical clinical symptoms, and approximately 40% of patients have already entered the terminal stage at the time of RC,^[3] missing the optimal treatment period. Moreover, the widely used tumor–node–metastasis (TNM) staging system cannot accurately evaluate the prognosis of patients with RC. If high-risk factors affecting the postoperative overall survival (OS) time of RC can be screened out early, targeted intervention, such as an intensive follow-up plan, can be implemented early, and the prognosis of patients can be improved, which has important theoretical and clinical significance.

Tumor depositions (TDs) are defined as isolated tumor foci within the perirectal fat without histologically recognizable lymph nodes or neural or vascular structures.^[4] However, the mechanisms underlying the formation and development of TDs remain unclear. Previous studies have shown that TDs status (present or absent) is an important prognostic indicator of RC.^[5] Overall survival and recurrence-free times were shorter in patients with TDs than in those without TDs. Tumors with TDs show strong aggressiveness and a high probability of distant metastasis. However, the impact of TDs count on the prognosis of patients was controversial. The newly proposed N1c category, specifically for patients with TDs but without lymph node metastasis, has been questioned by many scholars^[6,7] as the count of TDs was not considered and the actual risks posed by TDs were incompletely assessed.

Radiomics is a popular research method that utilizes high-throughput quantitative data extracted from medical images to reveal the physical, chemical, biological, and even genetic characteristics of tumors.^[8] The advantage of radiomics is that it is more objective than the traditional diagnostic method, which relies on the naked eye to observe the morphology and signal characteristics of lesions. To some extent, the radiomics method could solve the problem of diagnostic inconsistency between different diagnosticians. Radiomics has been widely used in disease diagnosis and therapeutic evaluation. For example, Yang et al developed an MRI-based radiomic model to predict the presence of lymph node metastasis,^[9] and Chen et al built a simulated neural-network-based ultrasound radiomic model to determine the presence of TDs.^[10] However, to the best of our knowledge, there is still a lack of evidence on whether the involvement of the TDs count in computed tomography (CT)-based radiomics models could further improve the prognostic prediction performance for patients with RC. Therefore, the aim of this study was to construct a nomogram based on clinicopathological factors (including TDs count), traditional CT manifestations, and CT radiomic features of RC to predict the prognosis of RC patients and to simultaneously investigate the impact of TDs count on the prognosis of RC.

2. Materials and methods

2.1. Patients

This retrospective study was approved by the institutional review board of West China Hospital. The requirement for written informed consent was waived. A total of 262 patients with rectal adenocarcinoma diagnosed between January 2013 and December 2018 (mean age 66.54 years; range 32–99 years) were recruited from West China Hospital of Sichuan University. The subjects underwent radical resection of RC in the Gastrointestinal Surgery Department of West China Hospital of Sichuan University, without receiving any radiotherapy or chemotherapy before surgery. The recruitment requirements were as follows: patients pathologically diagnosed with rectal adenocarcinoma; sufficient clinical information including patient’s age, sex, operation time, and post-operation survival information (dead or survival); patients who underwent laboratory tests (carbohydrate antigen19-9 and carbohydrate antigen 125) 1 week before surgery; and patients who underwent CT examination 1 week before surgery and CT images preserved intact in the PACS system. The exclusion criteria were as follows: patients who had not undergone CT examination preoperatively (−390), patients whose survival status was unclear (−520), CT images with motion artifacts or metal artifacts (−9), lack of preoperative laboratory test results (−24), and patients with other malignant tumors (−6). All patients enrolled in this study were randomly assigned to the training (n = 171) and validation (n = 91) groups at a ratio of 65:35. Figure 1 shows a flowchart of patient recruitment.

Figure 1.

Flowchart of patient enrollment and data collection.

2.2. CT examination protocol

In this study, portal venous phase CT contrast-enhanced images (iopromide injection [300 mgl/mL, 1 mL/kg]) were used to extract radiomics features. CT scanning was performed using the following types of scanners: Somatom Definite AS+, Siemens Healthcare Sector, Forchheim, Germany; Somatom Definite Flash, Siemens Healthcare Sector, Forchheim, Germany. The main parameters of the CT scanners are shown in Supplementary Table S1, Supplemental Digital Content, http://links.lww.com/MD/J234. Patients underwent special bowel preparations before the CT examination (they were advised to consume liquid food on the day before the examination and were administered a routine cleansing enema to ensure that the rectum was clear and empty 3–4 hours before the examination).

2.3. Reference standard for pathology

Pathological reports recording the pathological information of patients enrolled in this study were searched and downloaded from the Department of Pathology, West Chinese Hospital. In our study, pathological N, T stages, and diagnostic criteria for TDs were evaluated according to the American Joint Committee on Cancer 8th edition of the TNM staging system. The TDs status (positive or negative) was depicted clearly in the pathology reports. The pathological diagnostic criteria for TDs was defined as isolated tumor foci without histologically recognizable lymph nodes, neural, or vascular structures. The TDs count was documented in the pathology reports.

2.4. CT image evaluation

Two professional radiologists {with 10 years (radiologist Jin) and 15 years (radiologist Zhang) of experience in RC diagnosis}, who were both blinded to patient information, reviewed the CT images. Lesions classified as T1 and T2 stages on CT images were integrated to form the CT T1 to 2 cohorts. The mesoretal fascia was considered positive if the distance between the mesorectal fascia and the outermost edge of the tumor was <1 mm. Extramural vascular invasion was considered positive if dense blood vessels were present around the tumor. The tumor location was defined as follows: lower (<5 cm), middle (5–10 cm), and upper (>10 cm), based on the distance between the anal verge and lower pole of the tumor. Differences in opinion between radiologists were resolved through consultation during the image review. The baseline characteristics of the patients are shown in Table 1.

Table 1.

Baseline characteristics of study objects.

	Training group (n = 171)	Validation group (n = 91)	P value
Age	66.96 ± 12.08	65.75 ± 11.28	.418
Gender			.994
Male	103 (60.2%)	54 (59.3%)
Female	68 (39.8%)	37 (40.7%)
Vertical diameter (mm)	34.03 ± 12.79	35.89 ± 13.06	.271
Transverse diameter (mm)	15.68 ± 5.81	17.08 ± 8.03	.241
CEA			.604
(−, ≤5 ng/mL)	83 (48.5%)	48 (52.7%)
(+, >5 ng/mL)	88 (51.5%)	43 (47.3%)
CA19-9			.771
(−, ≤35 U/mL)	163 (95.3%)	86 (94.5%)
(+, >35 U/mL)	8 (4.7%)	5 (5.5%)
CA125			1.000
(−, ≤35 U/mL)	168 (98.2%)	90 (98.9%)
(+, >35 U/mL)	3 (1.8%)	1 (1.1%)
pT_stage			.487
T1	3 (1.8%)	3 (3.3%)
T2	58 (33.9%)	36 (39.6%)
T3	108 (63.2%)	50 (54.9%)
T4	2 (1.2%)	2 (2.2%)
pN_stage			.409
N0	107 (62.6%)	57 (62.6%)
N1a	21 (12.3%)	15 (16.5%)
N1b	21 (12.3%)	7 (7.7%)
N1c	12 (7%)	6 (6.6%)
N2a	9 (5.3%)	3 (3.3%)
N2b	1 (0.6%)	3 (3.3%)
Differentiation			.810
Lower	50 (29.2%)	30 (33%)
Middle	116 (67.8%)	59 (64.8%)
Untested	5 (2.9%)	2 (2.2%)
Microstate status			.554
Unstable	3 (1.8%)	0 (0%)
Stable	168 (98.2%)	91 (100%)
Ki-67	0.52 ± 0.17	0.5 ± 0.18	.868
TDs count {median (Q1–Q3)}	0 (0–0)	0 (0–0)	.219
MRF			.422
(−)	157 (91.8%)	81 (89%)
(+)	10 (5.8%)	5 (5.5%)
Untested	4 (2.3%)	5 (5.5%)
Location			.353
Lower(<5 cm)	80 (46.8%)	51 (56%)
Middle(≤5 to ≤ 10 cm)	76 (44.4%)	34 (37.4%)
Upper (>10 cm)	15 (8.8%)	6 (6.6%)
CT_stage			.433
T1~2	102 (59.6%)	54 (59.3%)
T3	66 (38.6%)	33 (36.3%)
T4	3 (1.8%)	4 (4.4%)
Perirectal_nodes (>3 mm)	1 (0~3)	1 (0~2)	.344

CA125 = carbohydrate antigen125, CA19-9 = carbohydrate antigen19-9, CEA = carcinoembryonic antigen, MRF = mesorectal fascia, TD = tumor deposition.

2.5. Follow-up

Follow-up began on the day of surgery and ended 3 years after surgery. Patients enrolled in this study were consistently monitored and followed up using chest X-ray and computed tomography at 3-month intervals during the first year, and at 6-month intervals in the following years.

2.6. Radiomics flowchart

2.6.1. CT image segmentation.

ITK-SNAP software (version 3.6.0; http://www.itksnap.org/) was used for image segmentation (Fig. 2). The region of interest (ROI) was manually drawn slice by slice within the border of the tumor regions by radiologist Jin with 10 years of experience in CT image diagnosis. All delineated 2-dimensional ROIs were automatically merged into a 3-dimensional volume of interest, and all volumes of interest were confirmed by the senior radiologist Zhang with >15 years of experience in RC diagnosis. Both radiologists were blinded to the pathological results. Any disagreement on the ROI border was resolved through consultation.

Figure 2.

Radiomics workflow.

2.6.2. Radiomic feature extraction and radiomics model building.

Radiomic features were extracted using Pyradiomics software (version 3.0.1; https://pyradiomics.readthedocs.io). CT images were resampled to an isotropic pixel spacing of 1.0 mm and the pixel intensities were rescaled to [0, 255] to correct the difference in pixel intensity and reconstruction algorithms caused by different scanners. CT images were preprocessed by image transformation {square, exponential, LBP2D and LBP3D, logarithm, square root, gradient, Laplacian of Gaussian filters, and low-pass and high-pass wavelet filters} to obtain more potential radiomic features that were most related to OS. The types of features extracted from the CT images included shape, first-order statistics, and texture features (gray-level size zone matrix, gray-level co-occurrence matrix, gray-level dependence matrix, and gray-level run-length matrix). Twenty patients were randomly selected from all study subjects to perform inter- and intra-class consistency (ICC) analyses. Jin and Zhang drew the ROI of the tumor to perform an interclass consistent analysis. Radiologist Zhang drew the ROI twice (1 week apart) to perform an intra-class consistent analysis. In the present study, the cutoff level of consistent analysis was set at 0.5 based on the following consensus: ICC 0.5 to 0.75, moderate consistency of features; ICC > 0.75, excellent consistency of features; and ICC < 0.5, poor consistency of features. Features with poor consistency will be deleted automatically, whereas those with ICC > 0.5 will remain and have a chance to be adopted to build the radiomic model. Pearson correlation coefficient analysis was used to remove redundant features, and the cutoff level of the correlation coefficient was set at 0.9. Then, least absolute shrinkage and selection operator (LASSO)-Cox analysis was then used to screen out features most associated with the OS of RC, and a 10-fold cross-validation method was applied to adjust and filter models to avoid model overfitting. Finally, the retained features and their LASSO-Cox regression coefficients were used to construct the optimal radiomics model.

2.6.3. Construction and evaluation of combined nomogram.

Univariate and multivariate Cox regression analyses were conducted to analyze the relationship between clinicopathological factors, traditional CT variables, and CT-based Rad-scores with patient prognosis (Table 2). Statistically significant factors (P < .05) in the multivariate Cox regression were used to construct the Cox regression model, which was then visualized using a nomogram. A time-dependent concordance index (C-index)curve was constructed to evaluate the predictive power of the model in the training and validation groups. A calibration curve was used to evaluate calibration of the nomogram. A decision curve analysis curve was used to evaluate the clinical usefulness of the prediction models.

Table 2.

Univariate and multivariate analysis of Cox regression.

Characteristics	Univariate analysis		Multivariate analysis
	Hazard ratio (95% CI)	P	Hazard ratio (95% CI)	P
Age	0.98 (0.93–1.04)	.517
Gender	1.47 (0.42–5.06)	.546
CA125	0 (0–Inf)	.998
CA19-9	2.32 (0.29–18.34)	.424
CEA	1.32 (0.37–4.68)	.669
Transverse diameter	0.98 (0.88–1.11)	.792
Vertical diameter	1.03 (0.99–1.08)	.16
pN_stage
N1a	1.66 (0.35–7.84)	.52
N1b	2.98 (0.77–11.51)	.114
N1c	1.42 (0.18–11.23)	.739
N2a	2.15 (0.27–16.97)	.468
N2b	0 (0–Inf)	.998
pT_stage
T2	0 (0–Inf)	.998
T3	5.23 (0.66–41.29)	.116
T4	20.29 (2.48–166.11)	.005	32.29 (3.31–315.49)	.0028
Differentiation
Lower	0.55 (0.12–2.62)	.456
Middle	1.22 (0.31–4.72)	.775
Untested	3.32 (0.42–26.21)	.256
TDs count	1.56 (1.09–2.24)	.016	1.6 (1.07–2.41)	.0237
Microstate status	9165855.8 (0–Inf)	.998
Ki-67	0.4 (0.01–19.36)	.646
Location
Higher	0 (0–Inf)	.998
Lower	0.76 (0.22–2.71)	.677
Middle	1.87 (0.53–6.63)	.332
MRF
(−)	0.77 (0.1–6.1)	.807
(+)	2.01 (0.25–15.85)	.509
Untested	0 (0–Inf)	.999
CT_stage	1.07 (0.3–3.81)	.911
T3	0.63 (0.16–2.45)	.508
T4	7.93 (0.99–63.22)	.051
Perirectal_nodes	0.86 (0.57–1.29)	.462
Rad_score	40.24 (1.65–981.38)	.023	74.12 (2.13–2575.39)	.0174

CA125 = carbohydrate antigen125, CA19-9 = carbohydrate antigen 19-9, CEA = carcinoembryonic antigen, MRF = mesorectal fascia, TD = tumor deposition.

“Inf” refers to the infinite.

2.7. Statistical analysis

Categorical variables were analyzed using the chi-square and Fisher exact tests. Numerical variables without normal distribution were analyzed using the Mann–Whitney U test, and numerical variables with normal distribution were analyzed by Student t test. Univariate and multivariate Cox proportional hazard regression analyses were conducted to select independent prognostic factors among clinicopathological factors, traditional CT manifestations, and Rad-scores. To measure the model performance, a time-dependent C-index curve was used and a value of >0.70 was considered as moderate diagnostic performance (C-index: 0.5–0.7, poor diagnosis efficient; 0.71–0.9, moderate diagnosis efficient; >0.91, excellent diagnosis efficiency^[11]). Kaplan–Meier survival analysis was used for OS risk stratification and comparison. For the decision curve analysis curves, the x- and y-axes represent the threshold probability and net benefit, respectively. If the prediction model has higher net benefits than the assumption that all patients die “ALL” or no patients die “NONE” or other models within the same range of threshold probability, it is considered to be clinically useful. Data were analyzed and plotted using the R software (version 4.2.0, http://www.r-project.org/). A 2-sided P < .05 was considered statistically significant.

3. Results

3.1. Patient’s baseline characteristics

A total of 262 patients were enrolled in this study, including 157 men, 105 women, 41 TD-positive, and 221 TD-negative patients. The mean age was 66.54 years, range, 32–99 years). The follow-up period was 36 months. The number of censored data points was 18 (6.87%). The 2- and 3-year survival rates were 86.64% (227/262) and 66.42% (174/262), respectively. The study subjects were randomly divided into a training group (n = 171) and a validation group (n = 91) at a ratio of 65:35. The characteristics of the variables between the 2 groups were even and there was no statistically significant difference. Table 1 provides a summary of patients’ baseline characteristics.

3.2. Selection of radiomic features and Radiomics model construction

For the consistency test of volumes of interest, total 2939 features were concluded in ICC test, and 832 features were removed. 2107 tumoral features had good reliability with ICC >0.5. Based on 10-fold cross-validation analysis, the optimal lambda value of LASSO-Cox regression was determined to be 0.04, which resulted in 4 radiomic features to construct the radiomic prognostic model (exponential_Firstorder_Kurtosis, gradient_firstorder_Minimum, original_shape_Maximum2DDiameterColumn, wavelet_HHL_ngtdm_Contrast) that performed best. As shown in Supplementary Figures S1–S3, Supplemental Digital Content, http://links.lww.com/MD/J235. The constructed optimal Rad-score was associated with OS with a hazard ratio (HR) of (HR = 74.12, 95% CI: 2.13–2575.39, P = .017) Supplementary File S1, Supplemental Digital Content, http://links.lww.com/MD/J236 shows the composition of the radiomics model.

3.3. Development and validation of radiomic-clinical combined OS prediction model

Three factors, namely Rad-score, pT_stage_T4, and TDs count, were independent risk factors for the OS prognosis of RC, and HRs (95% CI) were 74.12 (2.13–2575.39), 32.29 (3.31–315.49), 1.6 (1.07–2.41), respectively, with all P values of < 0.05. A Cox regression model composed of these independent risk factors was then developed (Table 2). A nomogram was used to visualize the constructed combined radiomic-clinical model (Fig. 3). The risk factors of OS in RC were divided into the high-risk and low-risk groups based on the median value of the Cox regression score, and the OS rate between 2 groups was compared, with HR (95% CI) of 9.25 (1.17–72.99) and log-rank of P = .01 (Supplementary Figure S4, Supplemental Digital Content, http://links.lww.com/MD/J237). Dynamic C-index values of the nomogram in 2 years were all >0.7 in the training group (Fig. 4A), and most of them were >0.7 in the validation group (Fig. 4B). The overall C-index of the nomogram model we built was 0.812 (95% CI: 0.707–0.917). The OS prediction performance of the nomogram model was compared with that of the clinical and radiomics models (0.812 vs 0.59, likelihood test of P = .019; 0.812 vs 0.714, likelihood test of P = .134) in the training group and (0.725 vs 0.585, likelihood test of P = .002; 0.725 vs 0.751, likelihood test of P = .256) in the validation group, respectively (Table 3). In the calibration curve, the x-axis represents the overall survival rate predicted by the nomogram model, the y-axis represents the actual overall survival rate observed, and the diagonal line indicates that the predicted value of the model is equal to the observed value. The red and purple lines represent the 2-year and 3-year overall survival rates predicted by the nomogram model, respectively. From the calibration curve, we found that the red and purple lines almost coincided with the diagonal lines, suggesting that the nomogram model had good diagnostic accuracy (Fig. 5A and B). Within the 0 to 1 threshold probability range, the nomogram model had a higher clinical net benefit and area under the decision curve in both the training and validation groups (Fig. 6A and B) than the clinical model and the radiomics model, respectively (within 0.125–0.3 threshold probability range, and the clinical net benefit of the nomogram was lower than that of the clinical model in the training group).

Figure 3.

Multi-parameter-based nomogram model. The risk factors are listed on the left side of the model. Each risk factor corresponds to a score by drawing a vertical line to horizontal axis, and the score of all risk factors is added to obtain an overall score, which corresponds to the 2-year and 3-year survival rates of rectal cancer patients.

Figure 4.

The time-dependent C-index curve for the nomogram predicting OS of rectal cancer patients. (A) In the training group. (B) In the validation group. C-index = concordance index, OS = overall survival.

Table 3.

Comparison of diagnosis efficacy of the model.

Group	Model	C_index (95% CI)	Likelihood value	P
Training group	Radiomics model	0.714 (0.581–0.848)	−47.78	.014
	Clinical model	0.59 (0.423–0.758)	−46.24	.019
	Nomogram model	0.812 (0.707–0.917)	−43.471	—
Validation group	Radiomics model	0.751 (0.58–0.922)	−34.28	.256
	Clinical model	0.585 (0.431–0.739)	−37.58	.002
	Nomogram model	0.725 (0.556–0.895)	−32.92	—

CI = confidence interval.

P refers to the comparison between Nomogram model and Clinical model/Rad_score model.

Figure 5.

The calibration curve for the nomogram predicting OS of rectal cancer patients. (A) In the training group. (B) In the validation group. OS = overall survival.

Figure 6.

The DCA curves of nomogram model, clinical model, and Radiomics model. (A) In the training group. (B) In the validation group. DCA = decision curve analysis.

4. Discussion

In this study, we developed a Cox regression model based on the clinicopathological factors and CT radiomic characteristics of RC, with the aim of identifying risk factors affecting the prognosis of RC and noninvasively evaluating the OS of RC. TDs count, pT _stage_T4, and CT-based Rad-score were independent risk factors for RC prognosis. The constructed Cox regression model performed better in predicting the survival rates of RC both in the training group (whole C-index [95% CI] = 0.812 [0.707–0.917]) and in the validation group (whole C-index [95% CI] = 0.725 [0.556–0.895]). The innovations of this study were first to build a Cox regression model involving TDs counts that had not been adopted in previous studies. Second, we visualized the model using a nomogram, which is a convenient tool for predicting prognosis. It can visualize complex mathematical equations and evaluate the prognostic risk of individual patients. In addition, we used a dynamic C-index curve to reflect the diagnostic efficiency of the model in real-time, which can describe the diagnostic efficiency of the model in more detail.

In the present study, we found that TDs count was an independent risk factor for OS in patients with RC. Previous studies have reported that TDs status (presence or absence of TDs) is an independent prognostic factor for patients with colorectal cancer. TDs are associated with poor prognosis in RC, advanced tumor stage, aggressive biological features, more intensive regional lymph node metastasis, and more perineural invasion.^[12] However, the impact of TDs count on the prognosis of patients was controversial, and only a few relevant studies exist. Author Song^[13] believed that considering TDs as positive lymph nodes in the pN category was potentially superior to the TNM (7th edition) pN category when evaluating the prognoses of RC patients, while Cohen^[14] found that the number of TDs had a linear negative effect on disease-free survival and OS. Qin^[15] demonstrated that the presence and number of TDs were both independent risk factors for the prognosis of patients with stage III colon cancer, and they found that the 5-year disease-free survival in patients with 1, 2 to 3, and ≥ 4 TDs count was 68%, 56.3%, and 0%, respectively (P < .001), and the 5-year OS was 76%, 59.4%, and 4.8%, respectively (P < .001). Similarly, our study also found that TDs count was an independent risk factor for prognosis of RC patients, and this finding added a variable to the personality assessment of prognosis of RC patients. In addition, our study provides evidence supporting the resolution of controversial N-category issues. Zhang^[16] reported that TDs’ number should not be ignored when making an accurate N stage, and they proposed a new N category method; for example, Zhang reported that the risk of death in patients with TDs > 2 was 2.051 times higher than that in patients with TDs ≤ 2 (95% CI: 1.191–3.531, P < .05), suggesting that a TDs count of 2 should be taken as the cutoff value, a TDs count <2 should be classified as N1c, and a TDs count >2 should be classified as N2c. Sholars Li^[17] proposed a TDs count of 3 as the cutoff value, for example, TDs count <3 should be classified as N1c, and TDs count ≥3 should be classified as N2c, and there have also been reports of supporting TDs counts of 4 as cutoff values.^[18] The category of N stage was controversial, and a reasonable N category method that reveals accurate risk of TDs count posed on RC patients’ prognosis needs to be verified by further large-sample studies. Our study not only contributed to a more comprehensive and accurate assessment of patient prognosis but also provided evidence supporting the resolution of controversial issues.

We also found that tumor pathological T stageT4 was an independent predictor of 2- and 3-year OS (HR = 32.29 [95% CI: 3.31–315.49]) of RC patients. The underlying mechanism may be that the higher the T stage of the tumor, the deeper the tumor invasion, the more aggressive the tumor behavior, the higher the probability of distant metastasis and local recurrence, and the worse the prognosis of patients. Kang^[19] also found that T stage was a risk factor for patient prognosis, and the tumor recurrence rate of patients with pathological T3 stage tumors was higher than that of patients with pathological T1 and T2 stage tumors (log-rank, P = .044). Moreover, in our study, tumor volume, rather than an independent risk factor, was a relevant factor for OS. Jiang^[20] revealed that tumor volume was an independent clinical indicator of OS; the OS rate of patients with a large tumor volume was lower than that of patients with a small tumor volume (80.4% vs 89.6%, P = .042). The reason for the inconsistency of study results with previous studies may be that the method {(slice thickness + slice spacing) × number of slices} we adopted to measure tumor volume caused a discrepancy between the real tumor volume and the results we measured.

As a data mining research method, radiomics has the advantage of revealing quantitative and objective research results. Its wide application in the clinic will be helpful in quantitatively explaining diseases and improving the accuracy of disease diagnosis. In our study, 4 radiomic features that were candidates for predicting the prognosis of patients with RC were screened. “Wavelet_HHL_ngtdm_Contrast” reflects difference between the pixel of the tumor lesion and the average gray value of normal lesion in the region. This means that the heterogeneity of the tumor, and as “Wavelet_HHL_ngtdm_Contrast” value increases, the value of the Rad-score and the risk of death of patients also increase. “Exponential_Firstorder_Kurtosis,” “Gradient_Firstorder_Minimum,” “Original_Shape__Maximum2DDiameterColumn” reflect the morphological characterize of pixel, with which relationship of OS of RC patients was shown by prognosis heat map.

Our study had several limitations. First, the small number of patients with TDs included in this study may have led to inaccurate results. Second, this was a retrospective study with inevitable selection bias. Third, only patients who underwent radical tumor resection were investigated, and the study results may be biased because patients who lost the chance of surgery in the advanced stages were not studied.

In conclusion, the multi-parameter-based Cox regression model involving TDs count and CT-based radiomics features is helpful for noninvasive evaluation of prognosis in patients with RC. The TDs count was an independent prognostic factor for RC.

Acknowledgments

There is no one to thank except members of our research team.

Author contributions

Formal analysis: Yumei Jin.

Methodology: Jun Zhang, Yewu Wang.

Software: Shengmei Liu, Ling Yang.

Supervision: Bing Song.

Validation: Hao Gu.

Writing – original draft: Yumei Jin, Siyun Liu.

Writing – review & editing: Siyun Liu.