Radiomics-clinical nomogram for preoperative lymph node metastasis prediction in esophageal carcinoma

Abstract

Objectives

This research aimed to develop a radiomics-clinical nomogram based on enhanced thin-section CT radiomics and clinical features for the purpose of predicting the presence or absence of metastasis in lymph nodes among patients with resectable esophageal squamous cell carcinoma (ESCC).

Methods

This study examined the data of 256 patients with ESCC, including 140 cases with lymph node metastasis. Clinical information was gathered for each case, and radiomics features were derived from thin-section contrast-enhanced CT with the help of a 3D slicer. To validate risk factors that are independent of the clinical and radiomics models, least absolute shrinkage and selection operator logistic regression analysis was used. A nomogram pattern was constructed based on the radiomics features and clinical characteristics. The receiver operating characteristic curve and Brier Score were used to evaluate the model's discriminatory ability, the calibration plot to evaluate the model's calibration, and the decision curve analysis to evaluate the model’s clinical utility. The confusion matrix was used to evaluate the applicability of the model. To evaluate the efficacy of the model, 1000 rounds of 5-fold cross-validation were conducted.

Results

The clinical model identified esophageal wall thickness and clinical T (cT) stage as independent risk factors, whereas the radiomics pattern was built based on 4 radiomics features chosen at random. Area under the curve (AUC) values of 0.684 and 0.701 are observed for the radiomics approach and clinical model, respectively. The AUC of nomogram combining radiomics and clinical features was 0.711. The calibration plot showed good agreement between the incidence of lymph node metastasis predicted by the nomogram and the actual probability of occurrence. The nomogram model displayed acceptable levels of performance. After 1000 rounds of 5-fold cross-validation, the AUC and Brier score had median values of 0.702 (IQR: 0.65, 7.49) and 0.21 (IQR: 0.20, 0.23), respectively. High-risk patients (risk point >110) were found to have an increased risk of lymph node metastasis [odds ratio (OR) = 5.15, 95% CI, 2.95-8.99] based on the risk categorization.

Conclusion

A successful preoperative prediction performance for metastasis to the lymph nodes among patients with ESCC was demonstrated by the nomogram that incorporated CT radiomics, wall thickness, and cT stage.

Advances in knowledge

This study demonstrates a novel radiomics-clinical nomogram for lymph node metastasis prediction in ESCC, which helps physicians determine lymph node status preoperatively.

Introduction

In accordance with the International Agency for Research on Cancer, the eighth most prevalent cancer in the world is esophageal cancer (EC), with 604 100 new cases in 2020 and an incidence rate of 3.1% of all tumours.¹ EC includes esophageal adenocarcinoma and esophageal squamous cell carcinoma (ESCC), with the former predominating in Western countries and the latter in China.² The incidence rate of EC in China is higher than the global average, and the number of cases accounts for more than half of all cases worldwide.³ In such a critical situation, we urgently need to improve the diagnosis and treatment of OC.

For OC in the thoracic segment without distant metastasis, surgery, radical chemoradiotherapy, and preoperative chemoradiotherapy combined with surgery are all important treatment tools. When a patient is diagnosed with OC, the assessment of lymphadenopathy status is critical to the choice of subsequent treatment modalities. Patients with lymph node-positive OC who receive preoperative chemoradiotherapy can have a survival benefit,⁴ and subsequently, preoperative chemoradiotherapy combined with surgery can be chosen as a treatment strategy for this group of patients. The choice of surgical approach may also differ for patients with and without mediastinal lymph node metastasis. Minimally invasive esophagectomy may have an advantage in lymph node dissection of 2 and 4 L mediastinum lymph node group, comparing with open esophagectomy.⁵

Current clinical modalities for the evaluation of lymph node metastases include ultrasound endoscopy, CT, MRI, and PET-CT. For some patients with esophageal strictures, ultrasound endoscopy may not be able to pass, and in addition it has a limited depth of detection and cannot detect lymph nodes far from the esophagus.⁶ Esophageal MRI is currently of limited use, and so far has been studied in small samples and has not been promoted in clinical practice.^7,8 When it comes to detecting regional lymph node metastasis in OC, PET/CT has a sensitivity that ranges from moderate to low, but it has a specificity that ranges from high to moderate. As a result, it has a certain false positive rate.⁹ The most widely used tool to determine lymph node metastasis in clinical practice is still CT, which is a comprehensive judgement based on lymph node location, size, and degree of enhancement. Nevertheless, the accuracy of the judgement is still limited.

The microscopic genetic or protein changes of tumours can be reflected in the imaging data, and we can better diagnose and predict the prognosis of tumours by extracting imaging features with high throughput.^10,11 Radiomics plays a significant part in the process of predicting postoperative lymph node metastasis in various malignancies including lung cancer, breast cancer, and gastric cancer.^12–14 Several previous studies based on CT radiomics to predict lymph node metastasis in ESCC have yielded preliminary results.^15–21 However, there is a lack of studies based on thin-layer contrast-enhanced CT, while the constructed prediction models lack information on the esophageal wall thickness and cT stage of ESCC. In the present study, we used thin-layer enhanced CT as the basis for the radiomics analysis, and the clinical features in the constructed model included the easily accessible clinical information of ESCC wall thickness and cT stage, so the present study compensated for the shortcomings of the previous studies to a certain extent. In summary, for the first time, we created a nomogram based on radiomics at thin-section contrast-enhanced CT (CECT) paired with wall thickness and cT stage to predict postoperative lymph node metastases in resectable ESCC.

Methods

Patients

This was a retrospective study and was approved by Medical Research Ethics Committee of Weifang People's Hospital (Ethical Review Approval No. KYLL20200921-1) exempting patients from informed consent. Patients who received radical surgical excision for EC at our hospital from January 2017 to December 2021 were included in this research. The following is a list of the primary criteria for inclusion in the study: (1) OC of the thoracic segment treated with radical surgery; (2) postoperative pathology of squamous cell carcinoma; (3) stage of I-IVa (American Joint Committee on Cancer, 8th edition); (4) preoperative availability of thin-section CECT of arterial phase; and (5) complete clinical and postoperative pathological data. The following list contains the primary criteria for exclusion: (1) postoperative pathology including adenocarcinoma, adenosquamous carcinoma, and small cell carcinoma and other pathological types; (2) cervical segment OC; (3) no preoperative thin-section CECT of arterial phase; (4) incomplete clinical and postoperative pathological data; (5) neoadjuvant therapy defined as preoperative treatment that encompasses chemotherapy alone, chemotherapy combined with immunotherapy, or chemoradiotherapy; And (6) no perceptible lesion on CT. Figure 1 depicts the study's flowchart.

Figure 1.

The study's flowchart, which illustrates the criteria for including and excluding patients.

CT examination

The CT scan was performed with the GE Discovery CT750 HD (GE, Milwaukee, United States), SOMATOM Drive (Siemens, Erlangen, Germany), and SOMATOM Force (Siemens, Erlangen, Germany). While the scan covers the chest or chest and abdomen, the patient's arms are raised and he or she breathes quietly. Specific CT scan parameters are shown in Table 1.

Table 1.

CT scan parameters.

Parameters	CT version
	SOMATOM drive (Siemens, Erlangen, Germany)	SOMATOM force (Siemens, Erlangen, Germany)	GE Discovery CT750 HD (GE, Milwaukee, United States)
CT tube voltage	120 kVp	120 kVp	120 kVp
CT tube current	4D care	4D	4D
CT rotation time	0.50 s	0.80 s	0.50 s
CT detector collimation	128 × 0.6 mm	192 × 0.6 mm	128 × 0.6 mm
Contrast agent type	Ioversol	Ioversol	Ioversol
Contrast agent dosage	1.5 mL/kg	1.5 mL/kg	1.5 mL/kg
Contrast agent infused rate	2.5-3.0 mL/s	2.5-3.0 mL/s	2.5-3.0 mL/s
Bolus-tracking threshold (aorta)	100 HU	100 HU	100 HU
Image matrix	512 × 512	512 × 512	512 × 512
Slice thickness	1 mm	1 mm	0.625 mm
Reconstruction interval	1 mm	1 mm	1.25 mm

Tumour segmentation

Figure 2 presents an illustration of the radiomics analysis workflow. Preoperative CT was retrieved from all enrolled patients, and the original Digital Imaging and Communications in Medicine (DICOM) images of 1 or 1.25 mm thin-section arterial phase CBCT were imported into 3D slicer (version 4.11.0, http://www.slicer.org), which is a free and open source image processing platform.²² One senior radiation oncologist with 7-year experience (X.G.) outlined regions of interest (ROI) of the 256 cases, which included primary OC. Generally speaking, the standard of OC lesion is the thickness of esophageal tube wall greater than 5 mm with obvious enhancement.²³ For some patients with stages I to II, when the thickness of the esophageal wall is less than 5 mm, we need to refer to the patient's barium X-ray radiography and esophago-gastro duodenoscopy (EGD) report to determine the extent of the esophageal lesion more precisely. Care was taken to avoid the esophageal lumen, esophageal contents, blood vessels, peri-esophageal fat, and artefacts when outlining. Thirty patients were randomly selected from all cases, and the ROI was outlined again by the same radiation oncologist 1 month later. A senior radiologist with 5-year experience (Y.Z.) also outlined the ROI of the 30 cases at the same time.

Figure 2.

Workflow of radiomics analysis.

Radiomics feature extraction

Preferred pre-processing is required before radiomics feature extraction, including resampling, greyscale discretization, and filter selection. The first step in the preprocessing procedure is resampling. Different voxels are a source of variability in the radiomics features of the images, and therefore the images need to be resampled to the same voxels.^24,25 Resampled voxel size was adopted to 1 mm × 1 mm × 1 mm according to previous study.²⁶ The second step of preprocessing is to adjust the greyscale. With a bandwidth of 25, the greyscale discretization of the images was applied.^26,27 The third part of the pre-processing is setting up the filter. We use Laplace of Gaussian (LoG) filter and wavelets filter. Based on previous research,²⁸ The LoG kernel sizes (sigmas) were 1.0, 2.0, 3.0, 4.0, and 5.0 mm, respectively, according to previous research. After pre-processing settings, the Radiomics extension pack in 3D slicer based on PyRadiomics platform was used for radiomics feature extraction.²⁹ A total of 1316 image radiomics features were extracted. The extracted imaging histological features included the following 8 categories: first-order features, shape (3D) features, shape (2D) features, grey level co-occurrence matrix features, grey level run length matrix features, gray level size zone matrix features, neighbouring grey tone difference matrix features, and gray level dependence matrix features. A detailed elaboration and calculation of the above radiomics features is provided in the official website of PyRadiomics (see https://pyradiomics.readthedocs.io/en/latest/features.html).

Radiomics and clinical features selection

We followed a 3-step procedure to identify radiomic features. First, we selected the radiomics features with good reproducibility by intra-class correlation coefficient (ICC). The intra-observer (radiation oncologist vs. radiation oncologist) and inter-observer (radiation oncologist vs. radiologist) consistency were calculated, and an ICC higher than 0.75 can be taken into account reproducible and can be used for further analysis.³⁰ Secondly, the selected radiomics features were normalized using Z-score standardization.³¹ Finally, we used LASSO logistic regression, with penalty parameter tuning conducted by 10-fold cross-validation, to select the most meaningful radiomics features. Instead of going through ICC test and Z-score standardization, the screening of clinical features was done directly through lasso regression, again with the coefficient with the smallest error selected through 10-fold cross-validation. The filtering criterion is the log (lambda, λ) value corresponding to the minimum mean square error plus its non-zero coefficient eigenvalue at one standard error.³² Selected radiomics features with non-zero coefficients were used to establish a radiomic score through a linear regression model.¹⁶

Development and assessment of the nomogram

Based on radiomics score and clinical characteristics, a nomogram was established to predict lymph node metastasis. The likelihood of lymph node metastasis is proportional to the total risk points calculated by adding the points for all variables. The nomogram's calibration was evaluated using calibration plots and Brier scores, and its discrimination was measured using the concordance index (C-index)³³ and receiver operating characteristic (ROC) curve. Decision curve analysis (DCA) was implemented in order to evaluate the clinical utility of the model.³⁴ In order to check the effectiveness of the model, 1000 iterations of 5-fold cross-validation were carried out.

Application of the nomogram

The total risk point for each patient was determined using the nomogram, and then patients were categorized as having either a low or high chance of developing lymph node metastasis. The optimized cut-off value of the risk point was calculated using the ROC, and this value served as the basis for identifying the criteria for grouping. Finally, we distinguished the high-risk group from the low-risk group by means of a confusion matrix^35,36 to further validate the predictive performance of the model for distinguishing between the presence and absence of lymph node metastases.

Statistical analysis

R software (version 4.1.0, available at http://www.rproject.org/) was utilized to carry out the statistical analysis for this study. The information is presented as means ± standard deviation (SD) of the data. When P was less than .05, statistical significance was assumed.

Results

Characteristics of the study cohorts

A total of 256 patients of ESCC were recruited in this study. The average age of the entire group was calculated to be 62.89 years; 225 patients were male; and 31 were female. 140 patients had lymph node metastasis, and 116 patients did not have lymph node metastasis. The clinical features include age, sex, smoking, drinking, tumour length, tumour location, wall thickness, cT stage, carcinoembryonic antigen (CEA), gross tumour volume (GTV), neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-monocyte ratios(LMR), systemic immune inflammation index (SII), systemic inflammation response index (SIRI), count of lymphocytes (LYMF), count of monocytes (MID), count of neutrophil granulocytes (GRAN), haematocrit (HCT), red blood cell distribution width (RDW-CV), platelet (PLT), and mean platelet volume (MPV). Platelet count times the neutrophil count divided by lymphocyte count is the formula for calculating SII. The equation for determining SIRI is the number of neutrophils multiplied by the number of monocytes and then divided by the total number of lymphocytes. The general characteristics of the cohorts are summarized in Table 2.

Table 2.

Characteristics of the study cohorts.

Characteristics	All patients (N = 256)
Age (years)	62.89 (7.74)
Sex (%)
male	225 (87.89)
female	31 (12.11)
Smoking (%)
No	76 (29.69)
Yes	180 (70.31)
Drinking (%)
No	80 (31.25)
Yes	176 (68.75)
Location (%)
Upper	7 (2.73)
Middle	56 (21.88)
Lower	193 (75.39)
Tumour length (cm)	4.72 (1.96)
Wall thickness (mm)	11.44 (4.51)
GTV (cm2)	17.33 (11.91)
Clinical T stage
T1-2 stage	72 (28.12)
T3-4 stage	184 (71.88)
Lymph node metastasis (%)
No	116 (45.31)
Yes	140 (54.69)
CEA (ng/mL)	2.46 (2.03)
NLR	2.56 (1.74)
PLR	150.28 (66.97)
LMR	4.53 (2.79)
SII	618.41 (406.15)
SIRI	1.24 (1.93)
LYMF (109/L)	1.79 (0.58)
MID (109/L)	0.45 (0.20)
GRAN (109/L)	4.10 (1.51)
HCT (%)	42.55 (4.44)
RDW_CV (%)	13.64 (15.69)
PLT (109/L)	243.71 (64.38)
MPV (fL)	10.14 (0.93)

Radiomics features extracted for constructing model

Four features of radiomics were extracted selected by the LASSO regression model including “original_shape_LeastAxisLength”, “log.sigma.4.0.mm.3D_glrlm_RunEntropy,” “wavelet.HLH_gldm_DependenceNonUniformityNormalized,” “wavelet.HHH_gldm_DependenceNonUniformityNormalized.” The process is shown in Figure 3.

Figure 3.

(A, B) Radiomics features were selected by the LASSO regression model.

Clinical features extracted for constructing model

The LASSO regression analysis extracted 2 clinical characteristics, including wall thickness and cT stage. The process is shown in Figure 4.

Figure 4.

(A, B) Clinical features were selected by the LASSO regression model.

Development and assessment of the nomogram

Based on the radiomics combined with clinical characteristics, we developed a nomogram model (Figure 5A). This model provided a quantitative representation of each variable and made it possible to perform individualized calculations of the patient's total risk point, which was correlated with the probability of lymph node metastasis. ROC curves and Brier scores were used for the evaluation of the discriminative ability of the model. The area under the curve (AUC) of the radiomics model and clinical model is 0.684 and 0.701, respectively. The AUC of the nomogram model including radiomics and clinical features is 0.711(Figure 5B). The model's calibration can be checked with the help of the calibration plot, which showed excellent agreement between predicted and actual likelihood of lymph node metastasis occurring (Figure 5C). The DCA demonstrated that for the same threshold probability, the nomogram yielded a greater net gain (Figure 5D). In order to conduct a more in-depth analysis of the performance of the model, we carried out a thousand-times 5-fold cross-validation on the AUC and the Brier score. The median value of the AUC was 0.702 (IQR: 0.65, 7.49), and the median Brier score was 0.21 (IQR: 0.20, 0.23) (Figure 5E).

Figure 5.

Presentation and assessment of the nomogram model. (A) presentation of the nomogram, (B) ROC curve of clinical model, radiomics model, radiomics and clinical model (nomogram model), (C) calibration plot, (D) DCA curve, € 1000 times 5-fold cross-validated AUC and Brier scores.

Utilization of the nomogram

To quantify the risk of lymph node metastasis in nomogram, we determined the sum of each patient's risk points and stratified them using the ROC curve's best cut-off score (Figure 6A). Risk point 110 (risk probability 54.14%) separated the low-risk group (35.5%, 91/256), from the high-risk group (64.5%, 165/256). Lymph node metastasis was 5.15 (95% CI, 2.95-8.99, P0.0001) times more common in the high-risk group compared to the low-risk group. For this particular risk stratification, the confusion matrix revealed an accuracy of 0.6914 (95%CI, 0.6309-0.7474) (Figure 6B).

Figure 6.

Utilization of the nomogram model. (A) The ROC curve provides the ideal cut-off value for risk stratification. [A total point was in below 110 was included in the low-risk group (without lymph node metastasis), otherwise the high-risk group (with lymph node metastasis)]. (B) Confusion matrix for evaluating disparities between anticipated and factual lymph node metastasis risk.

Discussion

In the present investigation, we found that the radiomics model combined with the clinical characteristics model could better predict postoperative nodal involvement in preoperative OC. The radiomics model includes 4 features, and the clinical model includes esophageal wall thickness and cT stage. Using radiomics features and clinical characteristics, we developed a nomogram of non-invasive method to assess lymph node metastatic status prior to surgery for OC.

The patient's general clinical profile, tumour markers, and haematological indicators are all examples of clinical characteristics. Based on reports in the past, we included CEA and several haematological signs that may be linked to lymph node metastasis. These haematological indicators included NLR, PLR, LMR, SII, SIRI, LYMF, MID, GRAN, HCT, RDW-CV, PLT, and MPV.^37–42 Two clinical factors, cT stage and tube wall thickness, were finally screened out after lasso regression. Yan et al³⁹ found that clinical T stage was closely associated with postoperative recurrent laryngeal nerve lymph node metastasis in OC by multivariate analysis of 430 cases of ESCC. Although there are no reports on wall thickness, wall thickness is closely associated with clinical T stage, and patients with large wall thickness are likely to present with late cT stage.

The presence or absence of lymphadenopathy involvement will affect the choice of treatment,⁵ and at the same time, lymph node metastasis can also be used as a prognostic indicator,⁴³ so it is crucial to determine lymph node metastasis before OC surgery. Judging whether lymph nodes are metastatic or not based on the size, morphology, and enhancement degree of metastatic lymph nodes of OC on CT images is a commonly used method in clinical practice, but the accuracy is limited, and there are problems of insufficient precision and specificity.^44,45 Ultrasound endoscopy is an invasive test to evaluate lymphatic specialties, but has potential complications such as anaesthesia risks, perforation of OC, and inability to cross esophageal strictures.⁴⁵ In recent years, in addition to CT and ultrasound endoscopy, the emergence of radiomics has provided new methods for preoperative determination of lymph node metastasis in OC. Radiomics captures the heterogeneity of different lesion tissues through high-throughput extraction of image features of lesion tissues. The biopsy tissue is small and can only represent the heterogeneity of the small piece of tissue that is biopsied, not the heterogeneity of the entire lesion tissue, which can be captured by the lesion radiomics features.⁴⁶ In comparison with other invasive tests and biopsies, radiomics is non-invasive, which greatly enhances its ease of use. In recent years, radiomics has developed rapidly and has been involved in many fields such as predicting postoperative T-stage, lymph node metastasis, lymphovascular invasion, prognosis, neoadjuvant radiotherapy efficacy, and radical radiotherapy efficacy in patients with OC.^47–49 Through a detailed review of the currently published articles on radiomics for the prediction of lymph node metastasis of ESCC,^15–21 I found the following shortcomings: First, the evaluation of the prediction models in some of the articles was incomplete or lacked evaluation of the models. Secondly, none of the previously published studies used confusion matrices, a tool for the assessment of model applicability. Thirdly, most of the articles had a CT layer thickness of 5 mm, and 1 article had a CT layer thickness of 1-1.5 mm in the external validation group, but the number of cases in the external validation group was relatively small, only 90 cases. Fourthly, most of the articles used only radiomics features for model construction, and only 2 articles used both imaging histological features and clinical features for model construction, but the clinical features included in the analysis were not comprehensive enough, and cT stage and esophageal wall thickness were not used in the previously published articles using clinical models. Compared with previous studies, this study has the following main merits: First, this study used a complete model assessment tool. After evaluating the model including ROC curve, Brier score, calibration degree curve, DCA curve, and a thousand-times 5-fold cross-validation, it was found that our model has good discriminating ability, calibration degree, and performance. Secondly, this study used a confusion matrix for the first time in a model for predicting lymph node metastasis of ESCC in radiomics-clinical nomogram to assess the predictive ability of the model. Thirdly, this study used all bitemporal thin-layer CECT, which may have made it more accurate in outlining esophageal lesions. Fourthly, this study included as many haematological indicators that are easily accessible in the clinic as possible when incorporating clinical features into the clinical features model, and although there was no statistical significance after lasso regression analysis, the inclusion of a comprehensive range of clinical features meant that we were more comprehensive and rigorous in selecting clinical imaging factors for the construction of the model. Moreover, this study is the first to incorporate cT stage and wall thickness, easily accessible clinical information, in the radiomics-clinical nomogram for predicting lymph node metastasis of ESCC.

In addition, our investigation has a number of restrictions. First of all, our study was a retrospective study from a single centre with a limited sample size. Due to the nature of retrospective studies, a proportion of patients without imaging data will be omitted, potentially affecting some useful information. Secondly, our study has not been validated by any other sources outside of our own research. It is likely that studies with external validation will be more convincing. The radiomics features that are screened by single-centre studies need to be externally validated to better determine the reproducibility of the selected features. Thirdly, because our study included patients for a relatively long period of time, the examination machines included 3 different types of machines, and the scanning parameters of each machine were not completely uniform.

Conclusions

In conclusion, our research indicates that quantitative radiomics features extracted from CT can accurately predict lymph node metastasis of ESCC prior to surgery. The radiomics-clinical nomogram, which integrated cT stage, wall thickness, and radiomics features, demonstrated superior predictive performance compared to the radiomics and clinical model. The nomogram based on CT radiomics and clinical features can help physicians determine the presence or absence of lymph node metastasis in patients with OC before surgery, thus achieving an accurate diagnosis. It is only under the premise of an accurate diagnosis that the physicians can select the most appropriate treatment plan for the patients.

Funding

This work was funded by the Science and Technology Development Project of Weifang City (Grant No. 2020YX008).

Conflicts of interest

None declared.