CT imaging-based radiomics signatures improve prognosis prediction in postoperative colorectal cancer

Abstract

OBJECTIVE:

To investigate the use of non-contrast-enhanced (NCE) and contrast-enhanced (CE) CT radiomics signatures (Rad-scores) as prognostic factors to help improve the prediction of the overall survival (OS) of postoperative colorectal cancer (CRC) patients.

METHODS:

A retrospective analysis was performed on 65 CRC patients who underwent surgical resection in our hospital as the training set, and 19 patient images retrieved from The Cancer Imaging Archive (TCIA) as the external validation set. In training, radiomics features were extracted from the preoperative NCE/CE-CT, then selected through 5-fold cross validation LASSO Cox method and used to construct Rad-scores. Models derived from Rad-scores and clinical factors were constructed and compared. Kaplan-Meier analyses were also used to compare the survival probability between the high- and low-risk Rad-score groups. Finally, a nomogram was developed to predict the OS.

RESULTS:

In training, a clinical model achieved a C-index of 0.796 (95% CI: 0.722–0.870), while clinical and two Rad-scores combined model performed the best, achieving a C-index of 0.821 (95% CI: 0.743–0.899). Furthermore, the models with the CE-CT Rad-score yielded slightly better performance than that of NCE-CT in training. For the combined model with CE-CT Rad-scores, a C-index of 0.818 (95% CI: 0.742–0.894) and 0.774 (95% CI: 0.556–0.992) were achieved in both the training and validation sets. Kaplan-Meier analysis demonstrated a significant difference in survival probability between the high- and low-risk groups. Finally, the areas under the receiver operating characteristics (ROC) curves for the model were 0.904, 0.777, and 0.843 for 1, 3, and 5-year survival, respectively.

CONCLUSION:

NCE-CT or CE-CT radiomics and clinical combined models can predict the OS for CRC patients, and both Rad-scores are recommended to be included when available.

Keywords

Colorectal cancer overall survival radiomics nomogram

1 Introduction

Colorectal cancer (CRC) is the third most common cancer worldwide, with a high mortality rate and poor prognosis [1]. The identification of prognostic factors in advance would be helpful in developing therapeutic strategies. Studies have indicated that clinical and pathological parameters could help predict survival in patients with CRC, including age, sex, tumor markers, pathological related factors, tumor-node-metastasis (TNM) stage, and liver metastasis, among others. The TNM staging system is recognised as one of the most important independent factors affecting the prognosis of patients with cancer [2]; however, the survival of patients with the same TNM stage still varies greatly because of the complexity of prognosis factors. Molecular biomarkers could be good prognostic factors for personalised treatment [3, 4], but these techniques are invasive, expensive, and not accessible to every patient [5]. Thus far, there is still no convenient and accurate method for predicting the prognosis of patients with CRC.

Radiomics, as a novel way to analyse diagnostic images, can identify the association with clinical outcomes through vast features extracted from medical images [5, 6]. Some studies on radiomics have also reported various clinical applications in colorectal cancer. Yang [7] investigated the feasibility of using a computed tomography (CT)-based radiomics signature to predict tumor genotype in CRC. Radiomic analysis of features extracted from fluorodeoxyglucose-positron emission tomography (FDG-PET)/CT [8] and magnetic resonance imaging (MRI) [9] has also been used to predict the survival of patients with CRC. CT is commonly used for lower abdominal symptoms owing to its popularity and convenience [10]. Researchers have used contrast-enhanced (NCE) or contrast-enhanced (CE) CT radiomics to predict colorectal cancer survival [11–13]. However, there is still uncertainty regarding whether NCE-CT or CE-CT is more accurate for predicting OS. Consequently, further research is required to determine the performance of these two modalities for OS prediction, as well as the potential performance improvement from combining them.

If preoperative NCE-CT and CE-CT based radiomics signatures (Rad-scores) could help improve prediction of OS for colorectal cancer patients, it would be helpful in developing individualized therapeutic strategies. In this study, the feasibility of Rad-scores from NCE-CT and CE-CT images before CRC resection surgery were investigated as prognostic factors for overall survival prediction of colorectal cancer patients.

2 Materials and methods

2.1 Patients

The study protocol was approved by an independent ethics committee review board of our hospital. Since this was a retrospective study, the requirement for written informed consent was waived for all participants. The workflow of this study is shown in Fig. 1. Data of patients who underwent CRC surgery between January 2010 and December 2011 at our hospital were retrospectively collected. This study included patients diagnosed with CRC who had complete follow-up data, pathological report information, and preoperative CT images, and had not received any preoperative therapy. Patients with mucinous adenocarcinoma were excluded. Ultimately, 65 patients were enrolled in the training set. Preoperative NCE- and CE-CT examinations one month before resection, postoperative pathological reports, age, sex, and OS data were collected. The eighth edition of the American Joint Committee on Cancer (AJCC) TNM classification was used for tumor staging [14]. According to the pathological TNM (pTNM) staging after surgery, observation was performed on patients in stage I-II. Among patients in stage III-IV, rectal cancer patients received combined radiotherapy and chemotherapy, while colon cancer patients received chemotherapy. Then the impact of different treatment approaches in the survivability for colon and rectal cancer patients in stage III-IV was then tested using log-rank test. The primary outcome of interest in our study was OS, which was interpreted as the period from diagnosis to death for any reason or the time of the last follow-up. The last follow-ups were performed in April 2017.

Fig. 1

Workflow of overall survival prediction model construction and validation.

For the external validation group, 28 cases were used from The Cancer Imaging Archive (TCIA), an open-source, open-access database of advanced medical imaging of cancer, which included 25 cases from The Cancer Genome Atlas (TCGA) colon adenocarcinoma (COAD) cohort [15] and three cases from rectum adenocarcinoma (READ) cohort [16]. Nineteen patients were finally included due to two cases with a loss of follow-up, two with mucinous adenocarcinoma, and five without CE-CT, as shown in Fig. 1. With the available diagnosis time and follow-up information provided in the database, survival status over time was calculated. The last follow-up was conducted from April 2011 to May 2015.

2.2 Image acquisition and segmentation

In the training set, all scans were obtained using two CT scanners (LightSpeed Ultra, GE Medical System, USA; Sensation 4, Siemens, Germany). The following scan parameters were used: 120 kV_P tube voltage, automatic mAs, 512×512 matrix size, and slice thickness of 5 mm for the abdomen and pelvic protocol or 7.5 mm for the chest to pelvic protocol. After unenhanced CT, the arterial and portal venous phases were performed at 35 and 65 s after contrast agent injection as a normal protocol. The contrast agent iohexol was injected at a rate of 2.5 ml/s. Portal venous phase CT images were chosen as the CE-CT images in this study.

The tumor area, as a region of interest (ROI), was drawn manually in both the CE-CT and NCE-CT images using the Elekta delineation workstation (Elekta, Sweden). The air and contents were excluded. Delineation was first performed by a junior radiation oncologist and then verified by a senior radiation oncologist to avoid feature differences caused by different delineations as much as possible.

For the validation set, the standard operation procedures (SOP) for these images were the same, with SOP Class UID 1.2.840.10008.5.1.4.1.1.2. The tube voltage and matrix size were the same as those in the training set. CE-CT images were collected before surgery, but NCE-CT images were not collected because they were not available for most patients. The segmentation procedure was the same as that used for the training set.

2.3 Radiomics feature extraction

For the training set, radiomics features of CE-CT and NCE-CT images with contoured target volumes were then extracted automatically using an in-house code with the Python package PyRadiomics, with modifying example CT parameters provided by the package. With a resampling method, CT images with different slice thicknesses and pixel spacings can be resampled with the same voxel spacing. Normalisation was performed because different scanners were used in this study. A total of 1046 radiomics features were extracted separately from NCE-CT and CE-CT images for each patient, including the shape, first order, Gray-level Co-occurrence Matrix (GLCM), Gray-Level Run Length Matrix (GLRLM), Gray-Level Size Zone Matrix (GLSZM) and Gray-level Dependence Matrix (GLDM) without a filter, and Laplacian of Gaussian (LoG) or wavelet filtering. In addition, the validation set followed the same procedure for CE-CT radiomics feature extraction as the training set, and 1046 features were extracted from the CE-CT images of each patient.

2.4 Feature selection and radiomics signature

The Z-score normalisation was performed for each radiomic feature. This feature normalisation can determine the influence of the large difference in absolute values between different features in the regression results. To reduce feature data redundancy, high-dimensional radiomics features were then reduced using correlation coefficients. Only one feature was accepted when the feature pairs had Spearman correlation coefficients >0.9. Univariate Cox regression was used to further screen the OS-related variables with p value <05. The least absolute shrinkage and selection operator (LASSO) method was subsequently used to further reduce feature dimensions. To ensure reliability of the Rad-scores, ten five-fold cross validations were performed for LASSO Cox regression to select the radiomics features that were associated with OS. Finally, one specific cross validation regression was chosen to construct NCE-CT and CE-CT Rad-score for each patient.

For each patient in the training set, the Rad-scores of the NCE and CE-CT images for each patient were calculated using the linear weighted summation of LASSO Cox selected features. For the validation set, the CE-CT Rad-score for each patient was constructed according to a regression algorithm based on the CE-CT image features selected in the training set.

2.5 Survival analysis and Nomogram

The log-rank test was firstly used to compare the differences in survival for different clinical variables. The clinical variables associated with OS were then investigated using univariate and multivariate Cox regression analyses. Combined Cox regression models derived from both the Rad-score and clinical variables were also constructed and compared. The model performance in the validation set was assessed using the C-index.

The threshold of the Rad-score was determined as the optimal cut-off value of the receiver operating characteristic (ROC) curve by maximizing the Youden index at the year time with the highest areas under the curve (AUC). The patients were then divided into high- and low-risk groups with the threshold in both the training and validation sets according to the NCE-CT and CE-CT Rad-scores, respectively. Kaplan-Meier survival analysis was used to assess the association between the radiomics signature and survival. A log-rank test was used to test the differences between the high- and low-risk groups.

Finally, a nomogram was developed to evaluate the survival of individual patients quantitatively. A calibration curve was plotted using a bootstrap of 1000 repetitions to assess the consistency between the observed and predicted survival probability of the nomogram. The survival prediction values generated by this prediction model enable the production of time-dependent ROC curves at 1, 3, and 5 year time points, which further evaluate the discrimination ability of the nomogram.

2.6 Statistics data analysis

Categorical variables were compared using the χ² test or Fisher’s exact test. The null hypothesis of the χ² test states that there is no significant difference between the observed and expected frequencies or proportions within the sampled data. Statistical analyses were conducted using R (version 4.0.5, http://www.R-project.org). Selection of radiomics features and logistic regression model (LASSO Cox) building were done using the “glmnet” package. The nomogram and its calibration curve were created using the “rms” and “survival” packages. Time-dependent ROC curve analysis was performed using the “survivalROC” package. Variables with p value <0.05 were included in a multivariate Cox regression model, and independent prognostic factors of survival were identified with p value <0.05. Kaplan-Meier curves were used to analyse the OS of the high- and low-risk groups, and a log-rank method was used to compare the differences. The Rad-score was calculated based on the regression coefficient of the selected radiomic features using multifactor linear weighting. For all tests, p value <0.05 was considered statistically significant.

3 Results

3.1 Clinical and pathologic characteristics

The training set comprised of 65 patients (37 male, aged 38–83 years) who underwent CRC resection surgery, with median and mean OS values of 56 and 50 months, respectively. The validation set comprised of 19 patients (9 male, aged 40–90 years), and their median and mean OS were 40 and 47 months, respectively. The clinical and pathological characteristics of patients are presented in Table 1. A log-rank test was used to test the difference of OS between the groups to investigate the correlations of clinical variables with OS. In the training set, the log-rank p values for age, N, M, TNM stage, and neurovascular invasion were less than.05. Furthermore, to assess the influence of different treatment approaches, log-rank p value of 0.282 was obtained in the training sets for colon and rectal cancer patients in stage III-IV.

Table 1
Clinicopathological characteristics and survival status of patients and their correlation with overall survival in both training and validation sets

Training set Validation set p (χ² test)

Variables No. of Patients 3-y OS 5-y OS Log-rank p No. of Patients 3-y OS 5-y OS Log-rank p

Age

≥65 33 19 9 <0.001 16 8 6 0.60 0.01

<65 32 26 22 3 2 2

Gender

Male 37 28 20 0.90 9 5 4 0.60 0.46

Female 28 17 11 10 5 4

Tumor location

Rectum 32 23 17 0.40 3 1 1 0.50 0.01

Colon 33 22 14 16 9 7

Pathology

pT stage

1-2 14 11 11 0.20 4 1 1 0.30 >0.99

3-4 51 34 20 15 9 7

pN stage

0 29 25 17 0.003 7 4 3 0.90 0.55

1-2 36 20 14 12 6 5

pM stage

0 62 44 31 0.002 16 8 6 0.60 0.25

1 3 1 0 3 2 2

PTNM stage

1-2 29 25 17 0.003 7 4 3 0.90 0.55

3-4 36 20 14 12 6 5

Histological grade

<3 54 40 28 0.07 –

3 11 5 3 –

Tumor size^*

≥4.75cm 35 24 17 0.80 –

<4.75cm 30 21 14 –

Tumor deposit

0 56 41 27 0.10 17 9 7 0.40 >0.99

1–3 9 4 4 2 1 1

Neurovascular invasion

Negative 57 42 30 0.02 12 6 5 0.04 0.03

Positive 8 3 1 7 4 3

	Training set	Validation set	p (χ² test)
Age
≥65	33	19	9	<0.001	16	8	6	0.60	0.01
<65	32	26	22		3	2	2
Gender
Male	37	28	20	0.90	9	5	4	0.60	0.46
Female	28	17	11		10	5	4
Tumor location
Rectum	32	23	17	0.40	3	1	1	0.50	0.01
Colon	33	22	14		16	9	7
Pathology
pT stage
1-2	14	11	11	0.20	4	1	1	0.30	>0.99
3-4	51	34	20		15	9	7
pN stage
0	29	25	17	0.003	7	4	3	0.90	0.55
1-2	36	20	14		12	6	5
pM stage
0	62	44	31	0.002	16	8	6	0.60	0.25
1	3	1	0		3	2	2
PTNM stage
1-2	29	25	17	0.003	7	4	3	0.90	0.55
3-4	36	20	14		12	6	5
Histological grade
<3	54	40	28	0.07	–
3	11	5	3		–
Tumor size^*
≥4.75cm	35	24	17	0.80	–
<4.75cm	30	21	14		–
Tumor deposit
0	56	41	27	0.10	17	9	7	0.40	>0.99
1–3	9	4	4		2	1	1
Neurovascular invasion
Negative	57	42	30	0.02	12	6	5	0.04	0.03
Positive	8	3	1		7	4	3

^*The patients were categorized into two groups based on the mean tumor size.

3.2 Feature selection and radiomics signature

The number of features was first reduced to 315 and 312 from the original 1046 for NCE-CT and CE-CT, respectively, using Spearman’s correlation. From the result of the ten times five-fold cross-validation LASSO Cox, the mean values of the C-index were 0.700 and 0.711 with relative standard errors of 1.61% and 0.64% for NCE-CT and CE-CT feature sets, respectively. A cross-validation LASSO Cox result was randomly chosen to construct the Rad-score. Table 2 lists the final selection of five features for NCE-CT and four features for CE-CT, determined through LASSO Cox.

Table 2
Radiomics features selected for NCE-CT and CE-CT

NCE-CT CE-CT

Original shape feature MinorAxisLength Maximum2DDiameterRow

Laplacian of Gaussian (LoG)

filtered feature Sigma.5.0. mm.3D_glcm_Idn –

Wavelet filtered features HLH_glszm_GrayLevelNonUniformity LHL_glszm_GrayLevelNonUniformity

HLH_glszm_ZoneEntropy LLH_firstorder_Kurtosis

LHH_glszm_ZoneEntropy LLL_glszm_ZonePercentage

	NCE-CT	CE-CT
Original shape feature	MinorAxisLength	Maximum2DDiameterRow
Laplacian of Gaussian (LoG)
filtered feature	Sigma.5.0. mm.3D_glcm_Idn	–
Wavelet filtered features	HLH_glszm_GrayLevelNonUniformity	LHL_glszm_GrayLevelNonUniformity
	HLH_glszm_ZoneEntropy	LLH_firstorder_Kurtosis
	LHH_glszm_ZoneEntropy	LLL_glszm_ZonePercentage

Significant differences were found (p = 0.02) between the two sets of C-indices for NCE-CT and CE-CT by applying the nonparametric Wilcoxon rank-sum test. The Rad-scores for NCE-CT and CE-CT were both calculated by a linear combination of the selected features and their coefficients for individual patients.

The resulting Rad-scores for NCE-CT and CE-CT were 0.105 (– 3.025, 2.355) and 0.144 (– 1.569, 1.290), respectively, for all patients in the training set, whereas the CE-CT Rad-score was – 0.125 (– 1.141, 1.212) for patients in the validation set.

3.3 Survival analysis and Nomogram

Table 3 shows the results of the univariate Cox analysis of the risk factors associated with OS. The results showed that the p values of age, N, M, TNM stage, and neurovascular invasion were all less than 0.05. Only age, TNM stage, and neurovascular invasion were included in the multivariate Cox regression analysis because TNM stage is a total stage with consideration of the T, N, and M stages.

Table 3
Univariate Cox regression analysis

Variable HR 95% CI p

Age at diagnose (≥65) 6.088 (2.089, 17.750) 0.001

Gender (Male) 0.931 (0.427, 2.030) 0.86

Tumor location (Rectum) 0.711 (0.327, 1.549) 0.39

T (3-4) 1.915 (0.656, 5.593) 0.24

N (1-2) 3.682 (1.476, 9.188) 0.01

M (1) 5.530 (1.625, 18.820) 0.01

TNM (3-4) 3.682 (1.476, 9.188) 0.01

Histological grade (3) 2.296 (0.919, 5.739) 0.08

Neurovascular invasion (Positive) 3.051 (1.131, 8.226) 0.03

Tumor size (≥4.75 cm) 1.089 (0.503, 2.359) 0.83

Tumor deposit (1–3) 2.042 (0.768, 5.429) 0.15

Variable	HR	95% CI	p
Age at diagnose (≥65)	6.088	(2.089, 17.750)	0.001
Gender (Male)	0.931	(0.427, 2.030)	0.86
Tumor location (Rectum)	0.711	(0.327, 1.549)	0.39
T (3-4)	1.915	(0.656, 5.593)	0.24
N (1-2)	3.682	(1.476, 9.188)	0.01
M (1)	5.530	(1.625, 18.820)	0.01
TNM (3-4)	3.682	(1.476, 9.188)	0.01
Histological grade (3)	2.296	(0.919, 5.739)	0.08
Neurovascular invasion (Positive)	3.051	(1.131, 8.226)	0.03
Tumor size (≥4.75 cm)	1.089	(0.503, 2.359)	0.83
Tumor deposit (1–3)	2.042	(0.768, 5.429)	0.15

Table 4 shows the results of the multivariate Cox regression analysis of the risk factors associated with OS and the model performance in the training set. Clinical and Rad-score combined models (Clinical-NCE model and Clinical-CE model) performed better than clinical or radiomics models alone. The two Rad-score combined model (Clinical-NCE-CE model) performed best for clinical use, with a C-index of 0.821 (95% CI: 0.743–0.899). In this model, age (HR: 5.577, 95% CI: 1.808–17.201) and neurovascular invasion (HR: 3.199, 95% CI: 1.091–9.379) were independent predictors for OS with p < 0.05, whereas for other factors, p≤0.10. The C-indices for Clinical-NCE model and Clinical-CE model were 0.807 (95% CI: (0.723–0.891) and 0.818 (95% CI: 0.742–0.894), respectively. The difference in performance among the three clinical radiomics combined models was small. Then, the Clinical-CE model was tested in the validation set. The C-index for the validation set was 0.774 (95% CI: 0.556–0.992).

Table 4

Multivariate Cox regression analysis for risk model associated with overall survival

		Model parameters		Model performance
Model name	HR	95% CI	p	C-index	95% CI
Radiomics model
NCE-CT	2.859	(1.632, 5.011)	<0.001	0.709	(0.599, 0.819)
CE-CT	3.056	(1.522, 6.137)	0.002	0.712	(0.602, 0.822)
Clinical model				0.796	(0.722, 0.870)
Age	7.778	(2.579, 23.463)	<0.001
TNM	3.920	(1.552, 9.896)	0.004
Neurovascular invasion	3.836	(1.332, 11.048)	0.01
Clinical-NCE model				0.807	(0.723, 0.891)
Age	5.292	(1.714, 16.344)	0.004
TNM	2.740	(1.022, 7.345)	0.045
Neurovascular invasion	3.576	(1.231, 10.393)	0.02
NCE-CT Rad-score	1.752	(0.993, 3.093)	0.053
Clinical-CE model				0.818	(0.742, 0.894)
Age	7.284	(2.420, 21.925)	0.00
TNM	3.240	(1.254, 8.374)	0.02
Neurovascular invasion	3.444	(1.193, 9.943)	0.02
CE-CT Rad-score	2.258	(1.022, 4.991)	0.04
Clinical-NCE-CE model				0.821	(0.743, 0.899)
Age	5.577	(1.808, 17.201)	0.003
TNM	2.480	(0.913, 6.737)	0.07
Neurovascular invasion	3.199	(1.091, 9.379)	0.03
NCE-CT Rad-score	1.612	(0.909, 2.859)	0.10
CE-CT Rad-score	2.035	(0.905, 4.572)	0.09

In the Cox regression analysis, the NCE-CT or CE-CT Rad-scores alone was an independent prognostic factor for OS (p < 0.05). Furthermore, cut-off values of 0.48 and 0.19 for NCE-CT and CE-CT, respectively, were obtained as the threshold of the ROC curve. Patients were further stratified using the Rad-score. In the training set, patients with Rad-score >0.48 were recognized as high-risk (N = 41) and those with Rad-score ≤0.48 as low-risk (N = 24) for NCE-CT; for CE-CT, those with a Rad-score >0.19 were categorised as high-risk (N = 28) and ≤0.19 as low-risk (N = 37). The distribution of Rad-scores for individual patients with high- and low-risk in both the training and validation sets is shown in Fig. 2.

Fig. 2

Radiomics scores for individual patient classified into high and low risk groups with a threshold of 0.48 and 0.19 for NCE-CT (A) and CE-CT (B) in the training set, and CE-CT (C) in the validation set.

In addition, the significant differences in survival were observed between patients in the low- and high-risk groups in both the training and validation sets according to the log-rank test (p < 0.05), as shown in Fig. 3.

Fig. 3

The Kaplan-Meier survival curves of the low- and high-risk groups according to the Log-rank test for patients in (A) training NCE-CT set, (B) training CE-CT set and (C) validation CE-CT set. Dotted lines are 95% confidence interval.

Finally, a nomogram was constructed using the combined regression model by integrating the clinical variables and CE-CT radiomics signature to predict the probability of the 1-, 3-, and 5-year OS after surgery for CRC, as shown in Fig. 4. It can be seen from the nomogram that age, TNM stage, and CE-CT Rad-score are all risk factors affecting OS. Age ≥65 years, TNM stage 3-4 and higher CE-CT Rad-score all led to an increased risk. The Rad-score had the longest line segment, indicating that it contributed the greatest risk to survival.

Fig. 4

A nomogram integrated with CE-CT radiomics signature and clinical variables. To use this nomogram, one should first locate the patient’s CE-CT Rad-score, age at diagnose, TNM stage and neurovascular invasion and draw lines straight up to the Points axis respectively to establish the score associated with that factor. Then adding up the scores of each factor to obtain the total score on the Total Points line. By drawing a vertical line intersecting with the 1-year, 3-year, and 5-year survival line, the corresponding survival probability for each patient could be determined.

Figure 5 shows the calibration curves for the evaluation of agreement between the model prediction and actual observation for the OS at 1, 3, and 5 years. The time-dependent ROC curve of the clinical and CE-CT Rad-score combined models is shown in Fig. 6. The areas under the curve were 0.904, 0.777, and 0.843 for survival predictions at 1, 3, and 5 years, respectively.

Fig. 5

The calibration curves of the nomogram for 1-, 3-, and 5-year survival probability.

Fig. 6

Time dependent ROC curves of the nomogram at 1, 3, 5-year period.

4 Discussion

In this study, the feasibility of preoperative NCE- or CE-CT image-based radiomics signatures was investigated as an independent prognostic factor for survival prediction in patients with CRC undergoing surgery. As can be seen from the results of log-rank test for clinical variables, age, N, M, TNM stage and neurovascular invasion were important prognosis factors that correlated with OS for postoperative CRC patients since their log-rank p values less than 0.05. From the training set, the performance of CE-CT radiomics features is slightly better than those of NCE-CT, and both are independent prognostic factors for OS. The Clinical-NCE-CE model performed slightly better than the other two combined models, with a C-index of 0.821 (95% CI: 0.743–0.899) in the training set for CRC survival prediction. A C-index of 0.818 (95% CI: 0.742–0.894) was achieved in the training set for the Clinical-CE model, with a value of 0.774 (95% CI: 0.556–0.992) in the validation set. In this model, factors including age, TNM stage, neurovascular invasion, and CE-CT Rad-score were associated with OS and were integrated into the final nomogram. Additionally, patients with stage I-II colorectal cancer only underwent surgical resection, while those in stage III-IV received different postoperative treatments. For the latter, no difference was found in OS despite different treatments (log-rank p value >0.05). Thus, the treatment factor was not included in the predictive model. Finally, AUCs of 0.904, 0.777, and 0.843 were achieved for the 1-, 3-, and 5-year survival prediction, respectively.

In the process of building the radiomics signature, the features were first selected by Spearman correlation and then by univariate Cox regression, resulting in relatively stable results that were ten times five-fold cross-validation LASSO Cox regression, with small relative deviations of C-index of only 1.61% and 0.64% for NCE-CT and CE-CT images, respectively. Although this process may be somewhat strict, the results were more stable.

Some differences were observed between the features selected from the NCE-CT and CE-CT images. LoG filter features are sensitive to areas with rapidly changing intensities, enhancing the edges, whereas wavelet filter features can more accurately describe the local signal features and separate the signal features from the background noise by considering the time-frequency variation (https://pyradiomics.readthedocs.io/en/latest/). These features can better reflect the essential content of the image; therefore, the selected features are mostly LoG or wavelet filters. Thus, we suggest including features with LoG or wavelet filters. In addition, as can be seen, LoG filter features are not selected for CE-CT, indicating that the LoG filter is more important for NCE-CT than it is for CE-CT. This is likely because the LoG filter enhances the image itself, making edge enhancement less important for CE-CT images.

Several other studies used preoperative NCE-CT images [11, 12], whereas others used CE-CT images [13] for OS prognoses. In this study, the radiomics signatures of NCE- and CE-CT images were both studied and compared as prognostic factors. It is evident that for both the radiomics alone model or the combined model, the CE-CT images were slightly better than NCE-CT images in terms of model performance. Furthermore, the model combining the two Rad-scores was slightly more advantageous. This is consistent with a report by Badic [17]. In their study, Badic [17] investigated the correlations between non-enhanced and enhanced CT scans in patients with CRC and suggested the extraction of complementary prognostic value from both NCE-CT and CE-CT modalities when available. Based on our findings, a combined model with the CE-CT Rad-score and clinical factors instead of NCE-CT is recommended for more accurate survival prediction in CRC patients. Additionally, it is highly recommended to include both types of images in the model if they are available, as also suggested by Badic’s study [17].

In this study, radiomics features extracted from preoperative CT images were associated with the survival of patients with CRC, as demonstrated by Rad-score HRs of 2.859 (95% CI: 1.632–5.011) and 3.056 (95% CI: 1.522–6.137) for NCE-CT and CE-CT, respectively, with p < 0.05. Similarly, Dai [12] demonstrated that the radiomics signature from pre-treatment NCE-CT can accurately predict OS with an HR of 3.053 (95% CI: 1.78–5.23; p < 0.001) for stage I– III colon cancer patients who underwent radical surgery, with an AUC of 0.768. Wang [11] also reported that radiomics features from radiotherapy treatment planning NCE-CT could predict OS with a model performance of 0.655 for patients with locally advanced rectal cancer.

With radiomics features alone, C-indices of 0.709 (95% CI: 0.599–0.819) and 0.712 (95% CI: 0.602–0.822) were achieved in the training set for NCE-CT and CE-CT radiomics models, respectively, in the prediction of survival for postoperative CRC patients in this study. This was slightly lower than that of the clinical model alone, with a C-index of 0.796 (95% CI: 0.722–0.870).

However, with the combination of radiomics features and clinical variables, the combined model constructed in this study outperformed the radiomics model alone in predicting OS (C-index 0.807 vs. 0.709 for Clinical-NCE model and 0.818 vs. 0.712 for Clinical-CE model). This was also reported in the study by Wang [11], where the model performance increased from 0.655 for a radiomics model to 0.730 for a combined model for locally advanced rectal cancer patients with radiotherapy planning NCE-CT. Xue [13] also reported a similar model performance of C-index which increased from 0.677 to 0.832 in an external test cohort for radiomics and clinical radiomics models for CRC patients with enhanced CT two weeks before surgery, with AUCs of 0.782 and 0.721 at three years for the training and internal validation cohorts, respectively. It is worth noting that although the Clinical-NCE-CE model had the best performance with a C-Index of 0.821, it only had a slight improvement compared to the clinical model’s C-Index of 0.796. This limited improvement may be attributed to the relatively small sample size, resulting in imperfect model construction. Nevertheless, this preliminary result suggests a trend in which combining Rad-score and clinical factors yields better performance compared to using standalone clinical or radiomics models.

The AUCs of our model were 0.904, 0.777, and 0.843 for the 1-, 3-, and 5-year survival, respectively. These results were comparable to those of Xue [13] and Li [18], with an AUC of 0.802 in the validation set for the three-year survival of rectal cancer patients who underwent curative surgery alone, with enhanced CT one month before resection. Additionally, in the comparison of AUCs, the 3-year survival AUC was found to be lower than the 5-year survival AUC, which could be attributed to the limited sample size. Despite the worst 3-year survival AUC being 0.777, it still suggests an acceptable overall performance.

This study also had some limitations. The first was its retrospective design. Although standardised image processing was performed, it was inevitable that there was still some heterogeneity due to the retrospective nature of the study. The second was the small sample sizes in the training and validation datasets, which may lead to data imbalance and affect the model performance. In future, larger sample size is still needed to validate and optimize the models for accurate and robust prognostic predictions and improved generalization capability. Thirdly, in the external validation set, there were only three cases involving the rectum, which may have introduced some bias. This validation set had a limited sample size and was obtained from different countries and regions. Factors such as race, diet, or genomic phenotype can significantly influence survivability. To address this, it is crucial to expand the availability of multi-center data in the future. Only the radiomics signatures of CE-CT images were tested in the external validation set, while the validation of the NCE-CT model still needs to be validated externally. Furthermore, the development of genetic testing technology has changed the treatment and prognosis of colorectal cancer [19]. In future work, the incorporation of genetic information is planned to further enhance the model performance.

5 Conclusions

The radiomics signature from preoperative CT images of patients with CRC undergoing resection surgery could be an independent prognostic factor for survival prediction. Enhanced CT images are recommended when available, whereas non-enhanced CT was also feasible with slight performance degradation. Furthermore, a nomogram combined with a radiomics signature and clinical variables was constructed to predict OS, which could be useful for developing treatment strategies.

Footnotes

Conflict of interests

The authors declare that they have no competing interests.

Funding

This work was supported by the National Natural Science Foundation of China (Grant No. 12205120), the Science and Technology Demonstration Project of Social Development of Jiangsu Province (Grant No. BE2019631), Double hundred Medical Professionals Program of Wuxi (Grant No. BJ2020053) and the Project of State Key Laboratory of Radiation Medicine and Protection, Soochow University (No. GZK1202202).

Acknowledgments

We sincerely thank the Information Centre for their help in retrieving CT images and the Department of Pathology for collecting pathological reports.

References

Sung

, Ferlay

, Siegel

et al., Global cancer statistics: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries, CA Cancer J for Clin 71(3) (2021), 209–249. doi: 10.3322/caac.21660

Goldstraw

, Chansky

, Crowley

et al., The IASLC lung cancer staging project: Proposals for revision of the TNM stage groupings in the forthcoming (Eighth) edition of the TNM classification for lung cancer, J Thorac Oncol 11(1) (2016), 39–51. doi: 10.1016/j.jtho.2015.09.009

de Cuba

E.M.V.

, Snaebjornsson

, Heideman

D.A.M.

, et al., Prognostic value of BRAF and KRAS mutation status in stage II and III microsatellite instable colon cancers, Int J Cancer 138(5) (2016), 1139–1145. doi: 10.1002/ijc.29855

, He

, Zhang

et al., A pair-wise meta-analysis highlights circular RNAs as potential biomarkers for colorectal cancer, BMC Cancer 19(1) (2019), 957. doi: 10.1186/s12885-019-6136-9

Horvat

, Bates

D.D.B.

, Petkovska

Novel imaging techniques of rectal cancer: What do radiomics and radiogenomics have to offer? A literature review, Abdom Radiol (NY) 44(11) (2019), 3764–3774. doi: 10.1007/s00261-019-02042-y

Gillies

R.J.

, Kinahan

P.E.

, Hricak

Radiomics: Images are more than pictures, they are data, Radiology 278(2) (2016), 563–577. doi: 10.1148/radiol.2015151169

Yang

, Dong

, Fang

et al., Can CT-based radiomics signature predict KRAS/NRAS/BRAF mutations in colorectal cancer? Eur Radiol 28(5) (2018), 2058–2067. doi: 10.1007/s00330-017-5146-8

Lovinfosse

, Polus

, Van Daele

, et al., FDG PET/CT radiomics for predicting the outcome of locally advanced rectal cancer, Eur J Nucl Med Mol Imaging 45(3) (2018), 365–375. doi: 10.1007/s00259-017-3855-5

Jalil

, Afaq

, Ganeshan

et al., Magnetic resonance based texture parameters as potential imaging biomarkers for predicting long-term survival in locally advanced rectal cancer treated by chemoradiotherapy, Colorectal Dis 19(4) (2017), 349–362. doi: 10.1111/codi.13496

10.

Ganguly

, Meredith

, Probert

et al., Colorectal cancer mimics: A review of the usual suspects with pathology correlation, Abdom Radiol 41(9) (2016), 1851–1866. doi: 10.1007/s00261-016-0771-2

11.

Wang

, Shen

, Zhong

et al., Radiomics features on radiotherapy treatment planning CT can predict patient survival in locally advanced rectal cancer patients, Sci Rep 9(1) (2019), 15346. doi: 10.1038/s41598-019-51629-4

12.

Dai

, Mo

, Han

et al., Prognostic and predictive value of radiomics signatures in stage I-III colon cancer, Clin Transl Med 10(1) (2020), 288–293. doi: 10.1002/ctm2.31

13.

Xue

, Peng

, Chen

et al., A CT-based radiomics nomogram in predicting the postoperative prognosis of colorectal cancer: A two-center study, Acad Radiol (2022)–10.1016/j.acra.2022.02.006

14.

Amin

, Edge

, Greene

et al., AJCC Cancer Staging Manual. 8th ed. In Springer; 2017.

15.

Kirk

, Lee

, Sadow

C.A.

et al., Radiology Data from The Cancer Genome Atlas Colon Adenocarcinoma [TCGA-COAD] collection, The Cancer Imaging Archive. http://doi.org/10.7937/K9/TCIA.2016.HJJHBOXZ

16.

Kirk

, Lee

, Sadow

C.A.

et al., Radiology Data from The Cancer Genome Atlas Rectum Adenocarcinoma [TCGA-READ] collection, The Cancer Imaging Archive. http://doi.org/10.7937/K9/TCIA.2016.F7PPNPNU.2016

17.

Badic

, Desseroit

M.C.

, Hatt

et al., Potential complementary value of noncontrast and contrast enhanced CT radiomics in colorectal cancers, Acad Radiol 26(4) (2019), 469–479. doi: 10.1016/j.acra.2018.06.004

18.

, Zhu

Y.Z.

, Zhang

Y.C.

et al. Radiomics of rectal cancer for predicting distant metastasis and overall survival, World J Gastroenterol 26(33) (2020), 5008–5021. doi: 10.3748/wjg.v26.i33.5008

19.

Benson

A.B.

, Venook

A.P.

, Al-Hawary

M.M.

et al., Colon cancer, version 2. NCCN clinical practice guidelines in oncology, J Natl Compr Canc Netw 19(3) (2021), 329–359. https://doi.org/10.6004/jnccn.2021.0012