Feasibility of an ADC-based radiomics model for predicting pelvic lymph node metastases in patients with stage IB–IIA cervical squamous cell carcinoma

Abstract

Objectives:

To investigate the prediction value of a radiomics model based on apparent diffusion coefficient (ADC) maps for pelvic lymph node metastasis (PLNM) in patients with stage IB–IIA cervical squamous cell carcinoma (CSCC).

Methods:

A total of 153 stage IB–IIA CSCC patients who underwent preoperative MRI including DWI from January 2015 to October 2017 were retrospectively studied and divided into a training cohort ( n = 102) and a validation cohort ( n = 51). Radiomics features were extracted from the ADC maps. The one-way ANOVA method, Mann-Whitney U test and Pearson’s correlation analysis were used for selecting radiomics features. Logistic regression analyses were used to develop the model. ROC analyses were used to evaluate the prediction performance of the model.

Results:

Clinical stage, tumor diameter, and MR-reported lymph node (LN) status were significantly associated with LN status ( p < 0.05 for both the training and validation cohorts). The radiomics model, which incorporated clinical stage, MR-reported LN status, and grey-level non-uniformity, showed good predictive performance in the training group (AUC 0.864; 95% CI, 0.782 – 0.924) and the validation group (AUC 0.870; 95% CI, 0.747 – 0.948). The performance of the radiomics model was significantly better than that of each predictive factor alone.

Conclusion:

The presented radiomics model, a non-invasive preoperative prediction tool, has the potential to have more predictive efficacy than clinical and radiological factors for differentiating between metastatic and non-metastatic lymph nodes.

Advances in knowledge:

A radiomics model derived from the ADC maps of primary lesions demonstrated good performance for predicting PLNM in stage IB-IIA CSCC patients and may help to improve clinical decision-making.

Introduction

Cervical cancer is the fourth most common malignant disease in females worldwide and the second leading cause of cancer death in young adults aged 20–39 years.^1–3 Although the evidence shows that the rate of cervical cancer is declining in developed countries,^4,5 the age-standardized incidence and mortality rates show a significant upward trend in females with cervical cancer in China.⁶ Cervical cancer, especially squamous cell carcinomas, which account for more than 80% of all cervical cancers, is a major contributor to the overall burden of disease in females worldwide.⁷ Radical hysterectomy and pelvic lymphadenectomy are the worldwide standard surgical options for stage IB–IIA cervical cancers, which is based on the International Federation of Gynecology and Obstetrics (FIGO) staging system that was recently revised in 2009.⁸ However, a large number of studies have shown that pelvic lymph node metastasis (PLNM) is found in less than 30% of patients with early-stage cervical cancer,⁹ which means that at least 70% of patients are over treated and have to bear complications of lymphadenectomy such as symptomatic postoperative lymphocysts and lower-limb lymphedema.¹⁰ In addition, it has been well-demonstrated that PLNM is an independent prognostic factor for recurrence and overall survival in early-stage cervical cancer.^11–13 Therefore, accurate knowledge of PLNM status before surgery is essential for patients with stage IB–IIA cervical squamous cell carcinoma (CSCC) to select the appropriate treatment methods and to adjust adjuvant treatments.

According to previous reports, some clinical and pathological features were regarded as significant risk factors of PLNM, such as clinical stage, tumor diameter, histological grade, parametrial invasion, and the depths of stromal invasion and lymphovascular invasion.^14–17 However, certain features, such as the depths of stromal invasion and lymphovascular invasion, are only assessable postoperatively. For this reason, some of the clinical and pathological features cannot be directly used in the preoperative prediction of PLNM.

Although the traditional size criterion for metastatic lymphadenopathy is a short diameter greater than 8 mm-10 mm,¹⁸ many research scholars have proven that this criterion cannot be considered completely satisfactory due to the existence of hyperplastic and micrometastatic lymph nodes (LNs).^19,20 Recently, studies have shown that diffusion-weighted imaging (DWI) is feasible for differentiating metastatic LNs from nonmetastatic LNs in patients with cervical cancer and that the apparent diffusion coefficient (ADC) value is an important parameter to identify the presence or absence of PLNM^20–22 ; however, the limitations are also very obvious. The main reason is that a node–by–node correspondence between DWI and histopathology could not be ensured, and hence there might not be identical results between the analysed LNs and the actual PLNM. The analysis of primary tumors in this study could avoid this problem to some extent.²³

Since the standardized concept of radiomics was proposed by Lambin in 2012,²⁴ radiomics data have been widely used in disease detection,²⁵ disease diagnosis,²⁶ evaluation of prognosis²⁷ and prediction of treatment response.²⁸ The principle of radiomics is extracting a large number of quantitative features from medical images using an automatic or semi-automatic high-throughput process, which effectively converts medical images into high-dimensional and mineable data to provide valuable information.^24,29 Currently, radiomics has been used in the prediction of LNs metastases in breast, colorectal, bladder and biliary tract cancer,^30–33 while PLNM in cervical cancer remains to be studied.

The aim of this study was to establish a preoperative prediction model for PLNM based on clinical-radiological factors and radiomics features of whole-lesion ADC maps according to the gold standard of post-surgical pathological findings and to explore the feasibility for differentiating metastatic LNs from nonmetastatic LNs in patients with stage IB-IIA CSCC.

Methods and Materials

Patients

The retrospective analysis was approved by the institutional review board of the hospital, and the required informed consent was obtained from all patients. A total of 153 patients with diagnosed stage IB–IIA CSCC who had surgical treatment between January 2015 and October 2017 were enrolled in the study. The inclusion criteria were as follows: 1) patients with a biopsy proven pathological diagnosis of CSCC, 2) patients with a clinical diagnosis of FIGO stage IB ~ IIA disease, 3) patients with no preoperative treatment and no other coexisting malignancies, 4) patients who underwent MRI including DWI examinations performed less than 20 days before surgery, 5) patients who underwent radical hysterectomy with pelvic lymph node dissection, and 6) patients with available clinical and pathological characteristics. The exclusion criteria were as follows: 1) patients with non-squamous cell carcinoma confirmed by surgical pathology, 2) patients with insufficient tumor size (maximal diameter <1 cm) for analysis, 3) patients who had scans exhibiting poor image quality or images with motion or susceptibility artefacts. Appendix F1 shows the patient recruitment pathway. The patients were divided into two independent groups according to a ratio of 2:1. The training cohort included 102 patients (mean age, 51.1 years; range, 27–70 years) who were recruited from January 2015 to April 2017, whereas the independent validation cohort of 51 patients (mean age, 50.4; range, 31–74 years) who were included from May 2017 to October 2017. The patient ages and clinical stages based on the FIGO staging system were obtained by reviewing the medical records.

MR examination

All MR examinations were performed using a 3.0 T unit (Magnetom Trio; Siemens Medical Solutions, Germany) with an 8-channel phased array body coil and respiratory gating technology. Patients were recommended to fast for 4–6 h before the examination. The MR scan covered the area from the superior edge of the iliac crest to the lower edge of the pubic symphysis.

Conventional MRI included transverse T ₁-weighted images (repetition time [TR] 514 ms, echo time [TE] 11 ms, slice thickness 5 mm, gap 2 mm, field of view[Fov] 512 × 640 mm, and number of excites [NEX] 2); transverse T ₂-weighted images (T2 fast spin echo, TR 3000 ms, TE 106 ms, slice thickness 5 mm, gap 2 mm, Fov 294 × 448 mm, and NEX 2) with fat suppression; and sagittal T ₂-weighted images (T2 fast spin echo, TR 3800 ms, TE 116 ms, slice thickness 4 mm, gap 0.8 mm, Fov 396 × 448 mm, and NEX 2). Transverse diffusion-weighted images were obtained using a single-shot spin-echo echo-planar imaging (EPI) sequence with the following parameters: TR 3500 ms; TE 93 ms; FOV 320 × 256 mm; slice thickness, 4 mm; gap, 0 mm; and b values of 0 and 1,000 s/mm² averaged in three orthogonal directions. The diffusion scan time was 91 s. A summary of the MRI parameters is presented in Table 1.

Table 1.

The summary table of MRI parameters

MR examination	Scanning parameter
	Imaging level	TR(ms)/TE (ms)	slice thickness (mm)	Gap (mm)	Fov (mm)	NEX
Conventional MRI	Transverse T1WI	514/11	5	2	512 × 640	2
	Transverse T2WI	3000/106	5	2	294 × 448	2
	Sagittal T2WI	3800/116	4	0.8	396 × 448	2
MR-DWI	Transverse	3500/93	4	0	256 × 320	2

Fov, field of view; NEX, number of excites; TE, echo time; TR, repetition time.

MR image acquisition and radiologic evaluation

Two radiologists reviewed all MR scans for consensus after evaluating the following traits for each CSCC patient: (a) tumor diameter, defined as the greatest tumor diameter measured on T ₂ weighted images in the sagittal, coronal, and transverse axial planes³⁴ ; and (b) presence of MR-reported LN-positive metastases, defined as larger than 10 mm in the maximal short-axis diameter that was measured on T ₂ weighted in transverse axial planes.³² The two radiologists were aware of the diagnosis of CSCC but were blinded to the clinical and pathologic details. Any disagreement was resolved through consultation.

Radiomics workflow

The radiomics workflow was as follows: (1) image segmentation; (2) radiomics feature extraction; (3) feature dimension reduction; and (4) prediction model development and validation.

Image segmentation

ADC maps that were exported by the DICOM protocol were imported into a prototype software programme (Omni Kinetics, GE Healthcare). Without knowing the results of the pathology examinations, a radiologist who had 7 years of experience used texture extraction software for 3D segmentation to form a volume of interest (VOI). The VOI consisted of regions of interest (ROIs) that were manually segmented along the tumor contour on each transverse section, with reference to T ₂ weighted and DWI (b = 0 s/mm²）images in the transverse axial planes. Each segmentation was validated by a senior radiologist who had 18 years of experience (largely with cervical cancer). ROI selection in patients with stage IB–IIA CSCC are presented in Figure 1. More information about the 3D segmentation procedure can be found in Appendix M1.

Figure 1.

ROI selection in patients with stage IB–IIA CSCC. ROI, region of interest.

Radiomics feature extraction

ADC maps were retrieved for radiomic feature extraction using the texture extraction software from Omni Kinetics. A total of 66 radiomics parameter features from ADC images were extracted for each patient, including first order features (n = 14), grey-level histogram features (n = 13), grey-level co-occurrence matrices (GLCMs) (n = 13), Haralick features (n = 10) and grey-level run-length matrices (GLRLMs) (n = 16). All features are calculated in three-dimensional directions within the whole-tumor volume. More information about the radiomics features can be found in Appendix M2.

Radiomics feature selection

The one-way analysis of variance (ANOVA) method and Mann-Whitney U test were used to select the useful predictive features among the 66 texture features in the training data set. To avoid a large number of feature variables affecting the predictive ability of the model or causing over fitting, feature redundancy was addressed using Pearson’s correlation analyses,³² which excluded the feature parameters with an auto-correlation coefficient higher than 0.7 (| r |>0.7).

Prediction model development and validation

Multivariable logistic regression analysis in a stepwise method was conducted to develop a model for predicting PLNM based on the clinical risk variables and the texture features of the ADC maps in the training cohort. The performance of the radiomics model was then internally tested in an independent validation set by using the formula derived from the training cohort.³³ Postoperative pathological findings were used as the gold standard for the training and validation of the model.

Statistical analysis

Receiver operating characteristic curve (ROC) analyses were performed using MedCalc statistical software version 15.2.2 (MedCalc Software bvba, Ostend, Belgium). Other analyses were performed using statistical software SPSS ver. 21.0 (IBM Co., Somers, NY). Univariate analysis was performed by comparing clinical-radiological characteristics between patients positive for PLNM and patients negative for PLNM using independent sample t-tests for numeric variables (e.g. age), and Chi-square test or Fisher exact test for categorical variables (e.g. clinical stage, tumor diameter and MR-reported LN status). Receiver operating characteristic curve (ROC) analyses were used to calculate the discrimination performances of both the radiomics model and the clinical-radiological risk factors in the training and validation datasets. P-values less than 0.05 were considered statistically significant.

Results

Clinical characteristics

Patient characteristics are summarized in Tables 2 and 3. There were no significant differences in the clinical and radiological characteristics between the training and validation cohorts. The rates of PLNM were 38.2% (39 of 102) and 29.4% (15 of 51) in the training and validation cohorts, respectively; however, no difference was found between the two cohorts (p = 0.282,>0.05). Therefore, the two cohorts can be used as training and validation cohorts. In both the training cohort of 102 patients and the validation cohort of 51 patients, there were statistically significant differences in the clinical stage, tumor diameter and MR-reported LN status (p < 0.05) between patients positive for PLNM and patients negative for PLNM, while there was no significant difference in age (p＞0.05).

Table 2.

Comparison of clinical-radiological features between training and validation Cohorts

Group	Number	Clinical stage		Tumor diameter		MR-reported LN status
		IB	IIA	≤4 cm	＞4 cm	LN-negative	LN-positive
Training cohort	102	71	31	63	39	86	16
Validation cohort	51	39	12	38	13	40	11
X2		0.792		2.462		0.810
p		0.373		0.117		0.368

LN, lymph node; X2, chi-square value.

Table 3.

Comparison of clinical-radiological features between LN-positive and LN-negative patients in the training and validation cohorts

Characteristic	Training cohort (n = 102)		p	Validation cohort (n = 51)		p
	PLNM(+)	PLNM (-)		PLNM (+)	PLNM (-)
Age, mean ± SD, years	50.18 ± 8.92	51.64 ± 9.06	0.430	49.47 ± 9.33	50.44 ± 9.73	0.959
Clinical stage			0.023 a			＜0.001 a
IB	22	49		6	33
IIA	17	14		9	3
Tumor diameter			0.001 a			＜0.001 a
≤4 cm	16	47		6	32
＞4 cm	23	16		9	4
MR-reported LN status			＜0.001 a			＜0.001 a
LN-negative	25	61		6	34
LN-positive	14	2		9	2

LN, lymph node; PLNM, pelvic lymph node metastasis.

p value is derived from the Chi-square test between each of the clinical variables and LN status.

^a p < 0.05.

Feature selection

Of the 66 extracted radiomic features, the four most stable features related to PLNM were considered for subsequent analysis and included: 1) max intensity, which was the greatest voxel-wise ADC values within the image region; 2) cluster prominence, which was a measure of the skewness and asymmetry of the GLCMs; 3) inverse difference moment, which was a measure of the local homogeneity of the GLCMs; and 4) grey-level non-uniformity, which was a measure of the similarity of the GLRLMs.

Establishment of the prediction model

Univariate analysis showed that the clinical stage, tumor diameter, MR-reported LN status and four texture feature predictors were associated with PLNM. Then, the above predictors were substituted into multivariate logistic regression analysis to establish a prediction model (Table 4). The model provided the estimated probability of PLNM for a particular patient with stage IB-IIA CSCC. This probability was equal to 1/ (1 + e^–z), where z = −3.509+(1.173 x clinical stage) + (3.191 x MR-reported LN status) + (0.168 x grey-level non-uniformity) , and e is the mathematical constant and base value of natural logarithms. According to the logistic principal, the critical value was 0.113, indicating that a probability of greater than 11.3% was taken to be diagnostic of PLNM.

Table 4. .

Multivariate analysis for predictive factors of PLNM

Predictive factors	B	SE	Wals	P	OR (95% CI)
Clinical stage	1.173	0.522	5.062	0.024	3.233
MR-reported LN status	3.191	0.844	14.282	0.000	24.303
Gray level non-uniformity	0.168	0.064	7.03	0.008	1.183
Constant	−3.509	0.869	16.299	0.000	0.030

95%CI, 95% confidence intervals; LN, lymph node; SE, standard error; Wals, Wald chi-square value.

B value is partial regression coefficient; OR value is Exp(B);

Independent validation of the radiomics model

All ROCs are provided in Figure 2a and b. In the training cohort, the radiomics model showed the highest discrimination between images with PLNM and images without PLNM with an area under curve (AUC) of 0.864 (95% CI, 0.782–0.924); the observed AUC value was higher than that of the clinical stage (AUC, 0.607; 95% CI: 0.505–0.702), tumor diameter (AUC, 0.668; 95% CI: 0.568–0.758) and MR-reported LN status (AUC, 0.664; 95% CI: 0.563–0.754). There were statistically significant differences among the AUC comparisons. In the validation cohort, the radiomics model yielded the greatest AUC value (0.870, 95% CI, 0.747–0.948), which confirmed that the radiomics model demonstrated better predictive efficacy than clinical stage (AUC, 0.744; 95% CI: 0.603–0.856), tumor diameter (AUC,0.744; 95% CI: 0.603–0.856) and MR-reported LN status(AUC, 0.772; 95% CI: 0.633–0.878). Table 5 presents the values of the radiomics model, including diagnostic accuracy, sensitivity, specificity, positive predictive value and negative predictive value.

Figure 2.

ROC curve analysis in the training and validation cohort.

Table 5.

The predictive efficacy of the radiomics model in the training and validation cohorts

	Accuracy	Sensitivity	Specificity	PPV	NPV
Training cohort (n = 102)	0.804	0.846	0.778	0.702	0.891
Validation cohort (n = 51)	0.784	0.867	0.750	0.591	0.931

NPV, negative predictive value; PPV, positive predictive value.

Discussion

A radiomics model based on the primary tumor characteristics was developed for the preoperative prediction of PLNM in stage IB–IIA CSCC. This prediction model incorporated three items, including the clinical stage, MR-reported LN status and one texture feature extracted from the ADC maps, and had an AUC value as high as 0.864 in the training set and 0.870 in the validation set. This study provides a new approach for the application of a radiomics model that has the potential to demonstrate more predictive efficacy than the clinical and radiological factors for differentiating between metastatic and non-metastatic lymph nodes.

Numerous scholars have proven that the risk factors of PLNM include clinical stage, tumor diameter, histological grade, parametrial invasion, LVSI and the depths of stromal invasion.^14–17 Based on some clinicopathological factors, Sun et al¹⁵ performed a retrospective analysis of 207 patients with early-stage CSCC and established a prediction model for PLNM using logistic regression analysis. The accuracy of the model was 76.3%, and the sensitivity and specificity were 53.8 and 83.9%, respectively. However, of these risk factors, only clinical stage and tumor diameter can be accurately obtained through medical records and imaging measurements before surgery. Although histological grade can be obtained from a presurgical biopsy, sampling bias with surgical tissue can lead to a high false-negative rate.³¹ The traditional size criterion of PLNM is a short diameter of less than 1 cm.¹⁸ However, these morphological evaluations with imaging techniques cannot be considered completely satisfactory, especially with very low sensitivity, which may be due to false negatives caused by small metastatic LNs (short diameter <10 mm) hidden in normal size LNs.¹⁹ Gong et al wrote about different imaging techniques for the detection of PLNM from gynaecological malignancies,³⁵ and the results of their meta-analysis show that the sensitivity of morphological evaluations with imaging is less than 50%.

Diffusion-weighted imaging, a non-invasive imaging method based on the properties of water diffusion, enables the observation of the molecular mobility of biological tissues.³⁶ A study demonstrated that ADC maps could accurately and clearly reveal the cellular density of primary cervical cancer tumors.³⁷ The study mainly analysed ADC maps and not the DWI data based on the fact that the selection of b-values has been demonstrated to significantly affect texture analysis on DWI images.³⁸ Moreover, one of the major challenges in radiomics is the process of managing feature redundancy to obtain a non-redundant set of features³⁹ ; this study reduced 66 texture features into four non-redundant features related to PLNM for subsequent analysis.

Radiomics has been proven to be able to noninvasively capture intra tumoural heterogeneity from routine medical imaging.^30,39 Since tumor heterogeneity has been related to tumor aggressiveness,⁴⁰ radiomics features of the primary tumor were selected to predict PLNM preoperatively in CSCC patients. In our study, the prediction model demonstrated better predictive efficacy than the clinical predictive factors alone, which indicates that the predictive model could be a better tool to predict PLNM in stage IB–IIA CSCC. In addition, the model also shows a higher predictive performance in the 51 patients of the validation group, with an AUC of 0.870. In the present study, the tumor diameter was related to risk of PLNM using univariate analyses, but the relationship did not show enough predictive strength by multiple logistic regression analysis so the variable was excluded for model development. The main consideration that can contribute to this controversy is that the rejection of tumor diameter predictors may be a result of the nuances in the data set or confoundment by other predictors.⁴¹ There is a correlation between tumor diameter and clinical stage, and PLNM has a strong correlation with MR-reported LN status, so the addition of tumor diameter might be not significant. In this regard, the rejection of tumor diameter does not definitively imply that it is unimportant. In our study, multiple logistic regression analyses showed that grey-level non-uniformity was included in the model, which represents the similarity of grey-values in the entire image.⁴² So the results meant that the more equal the grey-level distributions, the greater the possibility of PLNM in CSCC patients.

The limitations are as follows: First, the retrospective nature of the study may have a certain impact on the results, and the results need to be validated in a prospective study in further study. Furthermore, there is an obvious difference that the ADC values of squamous cell carcinoma were significantly lower than those of adenocarcinoma⁴³ ; this study only investigated CSCC to avoid the interactive effects of tumor types. Third, due to the small number of cases, it is necessary to expand the sample size to test the applicability of the radiomics model. Finally, associations between radiomics parameters based on DWI and other imaging modalities, such as PET/CT and dynamic contrast-enhanced MR, have not been investigated and require further research.

Conclusion

A key utility of this study is indicated by the result that the radiomics model has the potential to have more predictive efficacy than clinical and radiological factors for differentiating between malignant and benign pelvic lymph nodes, which could be helpful for patients with stage IB–IIA CSCC in selecting appropriate treatment methods and adjusting adjuvant treatments. Additionally, the results also help design a future prospective validation study that will be conducted to evaluate the performance of the model.