Multiparametric MRI radiomics nomogram for predicting lymph-vascular space invasion in early-stage cervical cancer

Abstract

Objective:

To develop a radiomics nomogram based on multiparametric MRI (mpMRI) to pre-operatively predict lymph-vascular space invasion (LVSI) in patients with early-stage cervical cancer.

Methods:

This retrospective study included 233 consecutive patients with Stage IB–IIB cervical cancer. According to the ratio of 2:1, 154 patients and 79 patients were randomly assigned to the primary and validation cohorts, respectively. Features with intraclass and interclass correlation coefficient (ICCs) greater than 0.75 were selected for radiomics features. The significant features for predicting LVSI were selected using the least absolute shrinkage and selection operator (LASSO) algorithm based on the primary cohort. The rad-score for each patient was constructed via a linear combination of selected features that were weighted by their respective coefficients. The radiomics nomogram was developed using multivariable logistic regression analysis by incorporating the rad-score and clinical risk factors.

Results:

A total of 19 radiomics features and 3 clinical risk factors were selected. The rad-score exhibited a good performance in discriminating LVSI with a C-index of 0.76 and 0.81 in the primary and validation cohorts, respectively. The radiomics nomogram also exhibited a good discriminating performance in two cohorts (C-index of 0.78 and 0.82). The calibration curve of the radiomics nomogram demonstrated no significant differences was found between prediction and observation outcomes for the probability of LVSI in two cohorts (p = 0.86 and 0.98, respectively). The decision curve analysis indicated that clinician and patients could benefit from the use of radiomics nomogram and rad-score.

Conclusion:

The nomogram and rad-score could be used conveniently and individually to predict LVSI in patients with early-stage cervical cancer and facilitate the treatment decision for clinician and patients.

Advances in knowledge:

The nomogram could pre-operatively predict LVSI in early-stage cervical cancer.

Introduction

The radical hysterectomy has been regarded as standard treatment strategy for patients with early-stage cervical cancer since 1895. ¹ Radical trachelectomy combined with lymphadenectomy is the most primary used fertility-sparing surgery in nulliparous patients with early-stage cervical cancer. ² The strategy to treat patients with early-stage cervical cancer requires a precise pre-operative assessment for clinical and pathological recurrent risk factors. ³

Previous studies have shown that lymph-vascular space invasion (LVSI) is an important independent prognostic factor for recurrence, overall survival, and disease-free survival in early-stage cervical cancer, especially in lymph-node-negative patients. ^4–8 LVSI was also a strong risk factor for lymph node metastasis (LNM) in patients with FIGO (International Federation of Gynecology and Obstetrics) Stage < IIB. ⁹ Furthermore, LVSI quantification of cervical histopathological sections could further add more useful and specified information for pelvic LN status. ¹⁰ Adjuvant radiotherapy or neoadjuvant chemotherapy is necessary in LVSI patients. Compared to LVSI patients without further treatment, LVSI patients with adjuvant radiation therapy had a 47% reduction in the risk of recurrence. ¹¹ If LVSI status could be identified pre-operatively, more tailored fertility-sparing treatment strategy could be adopted for nulliparous patients with early-stage cervical cancer. ^12–15

The presence of LVSI can be pre-operatively determined by specimens from cervical biopsy, cold-knife cone biopsy, and loop electrical excision procedures. However, Bidus et al indicated that specimens from these pre-operative cervical sampling techniques have poor positive-predictive values and/or low negative- predictive values for the determination of LVSI compared to the hysterectomy specimens. ¹¹ Up-to-date, there have been no effective pre-operative and non-invasive clinical or imaging approaches to predict LVSI in patients with cervical cancer. Therefore, it is critical to exploit a new method to predict LVSI status for cervical cancer, especially in its early stage.

Radiomics is the process of the conversion of medical images into high-dimensional, mineable data via high-throughput extraction of quantitative features, followed by subsequent data analysis for precision medicine. ^16–18 Liu et al reported that a radiomics nomogram enables a superior prediction of LVSI in breast cancer. ¹⁹ Based on a small sample size, recent radiomics nomogramstudies yielded fair or less satisfied performance for the prediction of LVSI in patient with cervical cancer. ^10,20,21 Therefore, this study aimed to develop a radiomics nomogram based on multiparametric MRI (mpMRI) and larger sample size to assess LVSI in patients with early-stage cervical cancer.

Methods and materials

Patients

The approval from the institutional review board was obtained, and the informed consent was waived in this study. From January 2016 to December 2018, 476 patients with histopathologically confirmed cervical cancer and pre-operative MRI scan were reviewed retrospectively. The inclusion criteria were as follows: (a) cervical cancer and LVSI status were pathologically confirmed based on hysterectomy specimens; (b) axial T ₁ weighted imaging (T ₁WI), fat-saturated T ₂ weighted imaging (FS-T ₂WI), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) maps and contrast-enhanced (CE) sequences were included in each MRI scan; (c) the interval between MRI scan and surgery was less than 30 days ²² ; and (d) the clinical data were available. The exclusion criteria were as follows: (a) pre-operative adjuvant radiotherapy or chemotherapy was performed; (b) poor imaging quality; and (c) tumor was invisible or less than three slices on MRI. Finally, 233 consecutive patients with early-stage cervical cancer were included in this study. The flowchart of the inclusion and exclusion criteria is shown in Figure 1. According to the ratio of 2:1, the 233 patients were randomly divided into the primary cohort (154 patients; mean age 50.1 ± 9.3 years) and the validation cohort (79 patients; mean age 49.9 ± 10.8 years). The clinicopathological characteristic of each patient were reviewed from the medical records, including age, tumor size and lymph node status on MRI (LNSM), FIGO stage, histological subtype, and Ki-67 level. LN positivity was reported from MRI if LN had one of the following characteristics: a circular shape, missing fatty hilum, signal inhomogeneity, or short-axis diameter larger than 10 mm. ^23,24

Figure 1.

The flow chart of inclusion and exclusion criteria in this study. ROI, region of interest.

MRI acquisition and segmentation

MRI was performed in a 1.5 T MR system (Magnetom Avanto, Siemens, Germany) with a phased-array body coil. MRI sequences and parameters are listed in Supplementary Table 1.

In the open-source imaging platform (MITK, v. 4.9.0; http://www.mitk.org), the region of interest (ROI) was manually delineated along the tumor contour on each axial FS-T ₂WI slice by a radiologist (the Reader 1, MLX, with 5 years of experience in pelvic imaging), then the ROI was automatically matched to T ₁WI, DWI (b factor of 1000 s/mm²), ADC map and CE sequences. The volume of interest (VOI) was constructed based on all slices of the ROI and feature extraction were performed. The mpMRI and the delineation of ROI on axial FS-T ₂WI are shown in Figure 2. The interclass and intraclass correlation coefficient (ICC) analyses are applied to assess the reproducibility of radiomics features. To evaluate the intraobserver and interobserver reliabilities, ROI delineation of 30 randomly chosen cases was repeatedly performed by Reader 1 within a 1 week interval and by another radiologist (Reader 2, YL, with 12 years of experience in pelvic imaging); both readers were aware of the diagnosis of cervical cancer but were blinded to the clinical and histopathological data. Features with intraclass or ICC greater than 0.75 were determined to have satisfactory reproducibility and were selected from radiomics feature extraction.

Figure 2.

A 42-year-old patient with cervical non-keratinizing squamous cell carcinoma, Stage IB2. The tumor invades the cervical stroma resulting in the thinner cervical stromal rim (arrow). (a, axial FS-T ₂WI; b, sagittal FS-T ₂WI; c, axial DWI; d, axial ADC maps; e, axial CE-T ₂WI; f, region of interest on axial FS-T ₂WI). ADC, apparent diffusion coefficient; CE-T ₁WI, contrast-enhanced T ₁ weighted imaging; DWI, diffusion-weighted imaging; FS, fat saturation; T ₂WI, T ₂ weighted imaging.

Radiomics feature extraction

The feature extraction was performed in Pyradiomics (v. 3.0.1; https://www.python.org). The feature calculation was based on primitive voxels and the feature were normalized by subtracting the mean value and then dividing by the standard deviation. To obtain better comparability of MRI gray values, a fixed bin width of 1 and 10 was used to compute original and wavelet-transformed features. Finally, the shape-based, first-order statistics, gray level cooccurence matrix (glcm), gray level size zone matrix (glszm), gray level run length matrix (glrlm) and gray level dependence matrix (gldm) feature matrices were calculated in original and wavelet-transformed forms. The detailed features names are listed in Supplementary Table 2.

Feature selection and Rad-score construction

The least absolute shrinkage and selection operator (LASSO) method was used in the primary cohort to obtain the most significant features for predicting LVSI in patients with early stage cervical cancer. In the process of feature selection, some unimportant features were eliminated by compressing their coefficients to zero and some significant features with nonzero coefficients were reserved that strongly associated with LVSI. The rad-score was calculated by the linear combination of selected features with their respective coefficients.

Construction of the radiomics nomogram

The clinical risk factors for LVSI entered in the radiomics nomogram were previously reported by other studies or identified by univariate and multivariate logistic regression analyses in this study. The radiomics nomogram was constructed by integrating the rad-score with the clinical risk factors using multivariable logistic regression analysis and was applied for individually predicting the probability of LVSI in the primary cohort.

Performance of the radiomics nomogram

The C-index, sensitivity and specificity were calculated to assess the prediction performance of the radiomics nomogram. To evaluate the calibration performance, the calibration curves were plotted using the Hosmer–Lemeshow (H–L) test which measured how close the prediction outcome generated by the nomogram was to the observation value. ²³ A significant result indicated a disagreement between the prediction value and the observation outcome of the nomogram.

Validation of the radiomics nomogram

The prediction performances of the rad-score and radiomics nomogram were tested in the validation cohort. According to the formula formed in the primary cohort, the total points were calculated for each patient in the validation cohort. The total points were also regarded as a factor when performed logistic regression analysis in the validation cohort. Finally, the C-index and calibration curve of the nomogram were derived in the validation cohort on the basis of the regression analysis and H–L test. The sensitivity and specificity of the radiomics nomogram were also calculated to assess the prediction performance in the validation cohort.

Clinical use

By quantifying the net benefits at different threshold probabilities in the validation data set, decision curve analysis was performed to determine the clinical usefulness of the radiomics nomogram. ²⁵

Statistical analysis

Statistical analyses were performed in R software (v. 3.0.1; http://www.Rproject.org), SPSS (v. 21.0.0.0; https://www.ibm.com) and MedCalc (v. 19.0.4; https://www.medcalc.org). A p-value less than 0.05 was considered statistically significant. After normality tested by the Shairo-Wilk test, continuous variables were expressed as the mean ± SD (normal distribution) or median and interquartile range (non-normal distribution), the differences in continuous or categorical variables between the LVSI and non-LVSI groups and between the primary and validation cohorts were compared using Student’s t-test, Mann–Whitney U test or χ² test, respectively. To further evaluate the association between LVSI and clinical risk factors, the univariate and multivariate logistic regression analyses were performed. The Delong test was conducted to compare the prediction performances of the rad-score and radiomics nomogram.

Results

Clinical characteristics

The comparisons of clinicopathological characteristics between patients with LVSI and those without LVSI and between the primary and validation cohorts are listed in Table 1. Among the 233 patients with early-stage cervical cancer, 151 patients with LVSI and 82 patients without LVSI. There was no significant difference in LVSI prevalence between the primary and validation cohorts (69% vs  60%, p = 0.08). There were no significant differences in the clinicopathological characteristics between the primary and validation cohorts or between the LVSI and non-LVSI groups (all P﹥0.05).

Table 1.

The comparisons of clinicopathological characteristics between the LVSI and non-LVSI groups, primary and validation cohorts

Characteristics	LVSI (+) (n = 151)	LVSI (-) (n = 82)	p-value	Primary cohort (n = 154)	Validation cohort (n = 79)	p-value
Age, mean ± SD, years	49.5 ± 10.1	50.9 ± 9.2	0.291	50.0 ± 9.3	49.9 ± 10.8	0.914
MRI-reported LN status			0.113			0.163
LN-positive	56 (37.1%)	22 (26.8%)		52 (33.8%)	26 (32.9%)
LN-negative	95 (62.9%)	60 (73.2%)		102 (66.2%)	53 (67.1%)
Tumor size (mm)	39.5 ± 11.6	39.7 ± 11.4	0.888	39.8 ± 11.8	39.1 ± 10.9	0.707
FIGO stage			0.751			0.664
IB1	74 (49.0%)	47 (57.3%)		79 (51.3%)	42 (53.2%)
IB2	27 (17.9%)	12 (14.6%)		26 (16.9%)	13 (16.5%)
IIA1	34 (22.5%)	16 (19.5%)		35 (22.7%)	15 (19.0%)
IIA2	15 (9.9%)	7 (8.5%)		14 (9.1%)	8 (10.1%)
IIB	1 (0.6%)	0 (0%)		0 (0%)	1 (1.2%)
Histologic subtype			0.312			0.211
Non-keratinizing SC	57 (37.7%)	37 (45.1%)		56 (36.4%)	38 (48.1%)
Keratinizing SC	69 (45.7%)	29 (35.4%)		70 (45.5%)	28 (35.4%)
Other subtypes	25 (16.6%)	16 (19.5%)		28 (18.2%)	13 (16.5%)
KI-67			0.107			0.211
≥50%	83 (55.0%)	54 (65.9%)		95 (61.7%)	42 (53.2%)
<50%	68 (45.0%)	28 (34.1%)		59 (38.3%)	37 (46.8%)

FIGO: 2009, International Federation of Gynecology and Obstetrics;LN: lymph node; LVSI: lymph-vascular space invasion; SC: squamous carcinoma.

Feature selection and rad-score construction

The 3940 features were extracted from the T ₁WI, FS-T ₂WI, CE, DWI, and ADC images of each patient. The 19 non-zero coefficient features (radiomics signature) were produced in feature selection using the LASSO method. The process of feature selection and the selected 19 features with their contribution coefficients are shown in Figures 3 and 4. The interobserver and intraobserver ICCs of 19 selected features ranged from 0.789 to 0.971 and from 0.781 to 0.985, which indicated that the 19 selected features have good inter- and intraobserver reliabilities and satisfactory reproducibility. The rad-score calculation formula was obtained by linearly combining these significant features with their respective coefficients (Appendix 3). Figure 5 shows the distributions of the LVSI status and rad-score in the primary and validation cohorts.

Figure 3.

The process of feature selection using the LASSO algorithm based on primary data. (a) The AUC curve is drawn against log(λ). The parameter λ(λ = 0.689) is chosen using 10-fold cross-validation via 1-SE criteria, and 19 features are selected by drawing vertical lines at the log (λ). (b) A coefficient profile plot is produced vs the log (λ). LASSO method compresses some feature coefficients to zero, and the features with non-zero coefficients, indicating a strong association with LVSI, are selected. A vertical line is plotted at the same optimal value resulted in 19 features with non-zero coefficients. AUC, area under the receiver operating characteristic; LASSO, least absolute shrinkage and selection operator; LVSI, lymph-vascular space invasion.

Figure 4.

Features and their respective contribution coefficients obtained from linear regression in process of constructing radiomics signature based on primary data. ADC, apparent diffusion coefficient; DWI, diffusion-weighted imaging; LVSI, lymph-vascular space invasion.

Figure 5.

Distributions of the rad-score and lymph-vascular space status in the primary cohort (a) and in the validation cohort (b). In patients rad-score greater than zero, the red bar represents true positive, the blue bar represents false positive. In patients rad-score less than zero, the blue bar represents true negative, the red bar represents false negative. LVSI, lymph-vascular space invasion.

Diagnostic validation of rad-score

Patients with LVSI had a rad-score of −1.56 to 2.58 (median 0.72), while patients without LNM had a rad-score of −3.23 to 1.17 (median 0.44). The rad-score exhibited good performance in discriminating between the LVSI and non-LVSI groups, with a C-index of 0.76 (95% CI, 0.68–0.84) and 0.81 (95% CI, 0.71–0.91) in the primary and validation cohorts, respectively. The rad-score yielded a sensitivity, specificity and AUC of 0.68 (95% CI 0.58–0.77), 0.75 (0.60–0.86) and 0.76 (0.69–0.83) in the primary cohort and 0.84 (0.71–0.94), 0.71 (0.53–0.85) and 0.81 (0.71–0.89) in the validation cohort, respectively, for predicting LVSI in early-stage cervical cancer (Table 2).

Table 2.

Diagnostic performance of rad-score and nomogram model in the primary and validation cohorts

Cohorts	AUC	SEN	SPE	+LR	-LR	PPV	NPV
Rad-score in primary cohort	0.76 (0.69–0.83)	0.68 (0.58–0.77)	0.75 (0.60–0.86)	2.72 (1.6–4.5)	0.43 (0.3–0.6)	0.86 (0.78–0.91)	0.51 (0.43–0.59)
Nomogram in primary cohort	0.78 (0.71–0.84)	0.90 (0.82–0.95)	0.58 (0.43–0.72)	2.15 (1.5–3.0)	0.18 (0.1–0.3)	0.83 (0.77–0.87)	0.72 (0.58–0.82)
Rad-score in validation cohort	0.81 (0.70–0.89)	0.84 (0.71–0.94)	0.71 (0.53–0.85)	2.87 (1.7–4.9)	0.22 (0.1–0.5)	0.79 (0.69–0.87)	0.77 (0.63–0.88)
Nomogram in validation cohort	0.82 (0.72–0.90)	0.80 (0.65–0.90)	0.77 (0.59–0.89)	3.40 (1.8–6.3)	0.26 (0.1–0.5)	0.82 (0.71–0.89)	0.74 (0.61–0.84)

AUC: area under the curve;+LR: positive likelihood ratio; -LR: negative likelihood ratio; NPV: negative-predictive value; PPV: positive-predictive value; SEN: sensitivity; SPE: specificity.

Development of an individualized prediction model

No significant characteristics for LVSI were identified through uni- and multivariate logistic regression analysis (Table 3). Previous studies indicated that age, tumor size and LN status were risk factors for LVSI in cervical cancer. Therefore, a radiomics nomogram was constructed by integrating the rad-score with age, tumor size and LNSM (Figure 6).

Table 3.

Univariate and multivariate logistic regression analyses for predictive factors of LVSI

Variables	Univariate analysis			Multivariate analysis
	Odds ratio	95% CI	p-value	Odds ratio	95% CI	p-value
Age	0.99	0.97–1.00	0.290	0.98	0.77–1.20	0.284
MRI-reported LN status	1.61	1.31–1.91	0.115	1.71	1.38–2.04	0.104
Tumor diameter (mm)	1.00	0.99–1.01	0.887	0.98	0.96–0.99	0.111
FIGO stage	1.16	1.02–1.29	0.273	1.24	1.09–1.38	0.145
Histologic subtype	1.09	0.90–1.28	0.658	1.05	0.85–1.25	0.821
Rad-score	1.87	1.66–0.27	0.030	1.96	1.75–2.17	0.002

FIGO, International Federation of Gynecology and Obstetrics.

Figure 6.

The radiomics nomogram is constructed by integrating rad-score with tumor diameter, patient age and LNSM (LN status on MRI) in the primary cohort. To use the nomogram, the scores are obtained by drawing lines on the point axis according to the patient age, tumor diameter, LNSM and rad-score, respectively. The total point was calculated by summing the scores of the four factors, then a vertical line to the bottom axis was plotted at the total point, which resulted in the estimated probability of LVSI in patient with cervical cancer. LNSM, LN status on MRI; LVSI, lymph-vascular space invasion.

LVSI prediction performance of the radiomics nomogram

The calibration curves of the radiomics nomogram for the probability of LVSI showed good agreement between the predicted and the observed outcomes in the primary and validation cohorts (Figure 7). The H–L test yielded no significant results in either cohort (p = 0.86 and 0.98, respectively), which suggested that no significant difference was found between the predicted outcomes and the observed outcomes. The radiomics nomogram also exhibited a good discriminating performance both in the primary and validation cohorts [C-index of 0.78 (95% CI, 0.70–0.86) and 0.82 (95% CI, 0.72–0.91), respectively]. The radiomics nomogram yielded a sensitivity, specificity and AUC of 0.90 (0.82–0.95), 0.58 (0.43–0.72) and 0.78 (0.70–0.86), respectively, in the primary cohort and 0.80 (0.65–0.90), 0.77 (0.59–0.89) and 0.82 (0.72–0.91), respectively, in the validation cohort for predicting LVSI in early-stage cervical cancer (Table 2). Although the prediction performances of the radiomics nomogram seems superior to those of the rad-score both in the primary and validation cohorts, there was no statistical significance (0.78 vs 0.76, p = 0.135; 0.82 vs 0.81, p = 0.737, respectively). The ROC curves of rad-score and radiomic nomogram for predicting LVSI in the primary and validation cohorts are in Figure 8.

Figure 7.

Calibration curve of the radiomics nomogram for predicting LVSI in the primary cohort (a) and in the validation cohort (b). No significant differences were found between the prediction and the observed outcomes of the radiomics nomogram (p = 0.86 and p = 0.98). LVSI, lymph-vascular space invasion.

Figure 8.

(a–d) are ROC curves of the rad-score and radiomic nomogram for predicting LVSI in the primary and validation cohorts, respectively. AUC, area under the curve; LVSI, lymph-vascular space invasion.

Clinical usefulness

The decision curve analysis showed that both the rad-score and nomogram could provide a net benefit to the patients (Figure 9). When the threshold probability was within a range from 1 to 54%, the net benefit of using the nomogram was better than that of either the treat-all scheme or the treat-none scheme.

Figure 9.

The decision curve shows that when the threshold probability from 1 to 54%, the radiomics nomogram adds more net benefit than schemes of treat-all and treat-none. LVSI, lymph-vascular space invasion.

Discussion

In this study, we constructed the models of rad-score and radiomics nomogram based on mpMRI to pre-operatively predict LVSI in patients with early-stage cervical cancer. The study demonstrated that both the rad-score and radiomics nomogram exhibited good performance in predicting LVSI, with C-index of 0.76 and 0.81 in the primary cohort, and 0.78 and 0.82 in the validation cohort and could be conveniently used to facilitate the individualized prediction of LVSI in patients with early-stage cervical cancer.

A total of 19 significant radiomics features were produced using the LASSO algorithm for data dimension reduction and feature selection from 3940 extracted features. Most features were wavelet features (except one original feature). The results indicated that wavelet-transformed features outperformed original features in predicting LVSI in patients with early-stage cervical cancer, and were consistent with previous study. ²⁶ The reason may be that the signal can be analyzed locally at any space domain in wavelet analysis, and the wavelet analysis has the ability to discover the structural characteristics hidden in the data that cannot be recognized by other signal analysis methods. ²⁷ Among 19 features, only 1 T ₂WI feature with small coefficient was selected. It indicated that T ₂WI sequence exhibited a limited ability in the models of rad-score model and radiomics nomogram.

The overfitting issue is extremely important in the process of constructing predictive model. We hold the opinion that possibility of overfitting bias was low in this study, the reasons are as follows: (1) the truth that the validation and primary cohorts yielded comparable performance is a reliable evidence to exclude the high possibility of overfitting bias; (2) although 3940 features were extracted, only 19 features were selected to construct the signature and nomogram. There were a total 106 patients with LVSI in the primary cohort, making an event-per-predictor ratio of 5.7. Vittinghoff proposed that the rule of thumb for event-per-predictor in logistic regression models should be between 5 and 9. ²⁸ Then, we think that overfitting bias was no big concern in this study.

Previous studies reported that. patient’s age, tumor size and LN status are risk factors for LVSI. Younger age, larger tumor and LNM are usually associated with a higher incidence of LVSI. ^29–34 Li et al indicated a significant difference in the incidence of LVSI between the tumor size >4 cm group and the tumor size ≤4 cm group (55% vs 34%). ²⁹ Zhang et al found that the incidence of LVSI was significantly different between patients aged ≤35 years and those aged >35 years (50% vs  39%). ³¹ Several studies have shown a significantly positive correlation between LNM and LVSI. ^10,32–34 However, this study showed no significant differences in patient’s age, tumor size and LN status on MRI between the LVSI and non-LVSI groups. And, the inclusion of the three factors to the radiomics nomogram did not significantly improve the prediction performances of the model. Compared to previous studies, the patients enrolled in this study were all less than Stage IIB except one case of IIB, which may compromise the difference between the LVSI and non-LVSI groups. Another reason maybe that MRI can’t reflect the true status of lymph node precisely.

Several studies reported the prediction performance of the radiomics and deep learning models for LVSI of cervical cancer, with the AUCs ranging from 0.73 to 0.91. ^12,20,21,35 Compared with these studies, our study had the following strengths: the model constructed in this study was based on mpMRI and more larger sample size; in the feature extraction, both wavelet and original features were calculated; The LASSO method was used to select features on the basis of the strength of association with LVSI. The radiomics nomogram had been confirmed to have the ability to generate an individualized probability of LVSI. In addition, we applied the decision curve analysis to calculate the net benefit and assess whether the model-assisted decisions would improve patient’s outcomes.

This study had some limitations. Firstly, it has been reported that LVSI in cervical cancer is highly correlated with DLL4 protein levels. ³⁶ The related serological biomarkers of cervical cancer, such as SAA, CA125, CA153, C199, and CEA, may have potential values in predicting LVSI in patients with cervical cancer. However, due to the incomplete information on these biomarkers, they were not included in the radiomics nomogram. Secondly, our study did not evaluate the value of the peritumoral tissue which may contain remarkable information about the development process of LVSI in early-stage cervical cancer. Thirdly, this study did not execute external validation of the model. It is necessary to carry out large-sample and multicenter prospective validation to provide compelling evidence for clinical application.

In conclusion, the radiomics nomogram developed by incorporating the rad-score with clinical risk factors can pre-operatively and non-invasively predict LVSI in patients with cervical cancer. Both the radiomics nomogram and the rad-score could be conveniently used to facilitate the individualized prediction of LVSI in patients with early-stage cervical cancer. The inclusion of clinical risk factors did not significantly improve the performance of the rad-score.