Background
According to current global cancer statistics, cervical cancer ranks second among female tumors, and the number of new cases in developing countries is increasing annually and involving younger women [
1]. Some studies have shown that LVSI is an independent prognostic factor in early-stage cervical cancer [
2‐
6]. LVSI was detected in 17.8% of early-stage cervical cancer cases in the initial pathological examination [
7]. Herr et al. [
3] found that the presence of satellite LVSI was associated with significantly reduced overall survival and disease-free survival, compared with lack of any LVSI. According to Memarzadeh et al. [
4], 86% of patients with perineural involvement in the parametria had evidence of parametrial LVSI, and the multivariate analysis revealed that large tumor size (> 4 cm), parametrial perineural invasion, cervical LVSI, and tumor depth (> 2/3) were significant simultaneous predictors of recurrence for early-stage cervical cancer (
p < 0.05). Subsequently, Pol et al. [
5] also confirmed that conjoined and satellite LVSIs were significantly associated with recurrence and survival. In a retrospective cohort study, LVSI was shown to be an independent factor that affects overall survival (
p = 0.009) and progression-free survival (
p = 0.006) in patients with early-stage cervical cancer [
6]. However, LVSI could only be confirmed by postoperative pathology.
In recent years, radiomics based on PET/CT imaging has been studied as an emerging tool for distinguishing tumor pathology stage, evaluating treatment response, and predicting prognosis [
8,
9]. Lambin et al. [
10] proposed the radiomics hypothesis that intratumoral heterogeneity evaluated by imaging could be the expression of genomic heterogeneity, as tumors with genomic heterogeneity are more likely to metastasize.
In previous studies, some protein molecules expression have been shown to be associated with lymphangiogenesis, lymph node metastasis [
11], and lymphovascular invasion in the early-stage cervical cancer. Hoellen et al. [
12] determined that cyclooxygenase-2 (COX-2) expression was significantly associated with LVSI (
p = 0.017). The expression of tenascin-C (TNC) in invasive cervical carcinoma was markedly increased [
13]. Pilch [
13] et al’s research showed that, in 84% of the cases examined, a strong TNC immunoreactivity was noted around and within the tumor cell nests. TNC is not only associated with epithelial–mesenchymal transition, proliferation, and migration of cancer cells, but it also facilitates the formation of cancer stroma, including desmoplasia and angiogenesis [
14]. Thus, we aimed to explore the association of radiomics derived from
18F-fluorodeoxyglucose PET/CT imaging combined with COX-2 and TNC expressions with LVSI in early cervical cancer and to establish a machine learning model of the combination of PET/CT radiomics, COX-2, and TNC to predict LVSI in patients with early-stage cancer.
Patients and methods
Patient cohort
This monocentric and retrospective study was performed at the Department of Radiology at Shengjing Hospital in Shenyang, China. Between January 2015 and December 2019, 131 female patients with cervical cancer confirmed by biopsy pathology underwent pretreatment
18F-FDG PET/CT. Inclusion criteria: (1) histologically confirmed cervical cancer with stage Ia–IIa determined by the 2018 International Federation of Gynecology and Obstetrics (FIGO) classifications [
15]; (2) absence of other malignant tumors; (3) normal serum glucose levels before undergoing PET/CT. Of these patients, 19 subjects were excluded because of previous chemoradiotherapy before the examination (
n = 5), tumor volume < 1 cm
3 leading to image data being unsuitable for textural feature measurement (
n = 10), and surgery performed in another hospital (
n = 4). Finally, 86 patients (42 LVSI and 44 non-LVSI) were randomly divided into two groups, including the training and testing datasets, according to a 7:3 ratio. An additional 26 patients were used as an external dataset for model validation.
Immunohistochemistry
All tissue specimens of cervical carcinoma were prepared by the Department of Pathology in our hospital. Immunohistochemical staining was performed using Leica BOND MAX™ (Leica Biosystems, Shanghai, China). For immunohistochemical detection of COX-2 and TNC protein expressions, the sections were incubated with goat anti-human COX-2 (1:400 dilution) or rabbit anti-human TNC (1:400 dilution) polyclonal primary antibodies (both from Abcam, Shanghai, China) at 4 °C overnight. After washing, the sections were incubated with species-appropriate enzyme-conjugated anti-rabbit and anti-goat secondary antibodies. Dewaxing, antigen epitope exposure, blocking, incubation with primary antibody, development of diaminobenzidine oxidation color, hematoxylin staining, and dehydration were automatically completed by the computer.
After the tissue was sliced, it was placed on Pannoramic MIDI tissue slice scanner (3DHISTECH Ltd., Budapest, Hungary), which simultaneously moved and scanned the image, forming a file that contained all the information on the tissue section. The file could be magnified 1–400 times using the Pannoramic viewer software, and the picture could be intercepted at any location. The QuantCenter (3DHISTECH Ltd.) is an analysis software that supports the Pannoramic viewer. After image scanning is completed, the DensitoQuant software in the QuantCenter automatically recognizes and sets all dark brown areas on the tissue section as strong positive, brown yellow as moderate positive, light yellow as weak positive, and blue cell nucleus as negative. Furthermore, all strong-positive, moderate-positive, weak-positive, and negative areas (in pixels); the percentage of positive areas; and the H-score (Immunohistochemical score) were analyzed for each tissue [
16]. According to the previous study, H-score = staining intensity [negative (0), mild (+ 1), moderate (+ 2), intense (+ 3)] * percentage (%) of positive stained cells [negative (0), < 5% positive stained cells (+ 1), 5–20% cells positivity (+ 2), 21–50% cells positivity (+ 3), (> 50% cells positivity (+ 4)]. All cases were divided into three groups according to the extent of staining and H-score: negative or weak-positive(H-score, 1 and 2), moderate-positive(H- score,3,4 and 6), and strong-positive groups(H-score,8,9 and 12) [
17].
18F-FDG PET/CT acquisition and features extraction
The patients rested quietly for 60 min before PET/CT (Discovery PET/CT 690; GE Healthcare, Chicago, IL, USA) scanning. All CT and PET scans were acquired with free breathing for attenuation correction and image fusion. First, low-dose non-enhanced CT images were acquired with a bulb voltage of 120 kV, auto mA (30–210 mA; noise index, 25), and slice thickness of 3.27 mm. Then, the PET data were acquired after CT scanning using a three-dimensional acquisition mode at a speed of 1.5 min/bed (7–8 beds in total) and a matrix size of 192 × 192. The time-of-flight and point-spread function techniques were also used in the reconstruction. A volume of interest (VOI) of tumor was automatically obtained on an AW4.5 workstation (GE Healthcare) using a threshold of 42%SUVmax. The metabolic tumor volume (MTV), total lesion glycolysis (TLG), and maximum (SUVmax), mean (SUVmean), and peak (SUVpeak) standard uptake values were measured automatically corrected with body weight inside the segmented VOI.
The VOI based on PET images was manually drawn layer by layer around the cervical squamous cell carcinomas by two nuclear medicine physicians, who performed the task independently. Both had more than 10 years of experience and were blinded to patients’ clinical data. In order to ensure the stability of segmentation, every nuclear medicine physician independently segmented VOI of all images. Then, 40 VOIs of PET image from two doctors were randomly selected to extract radiomics features and calculate intraclass correlation coefficient (ICC). In the end, the VOIs of one doctor were randomly selected to be used for research. The radiomics features were extracted from the PET images using AK software (Artificial Intelligence Kit, GE Healthcare, Shanghai, China) [
18].
Calculation of Radiomics score
All data analyses were processed with R version 3.5.1 (The R Foundation, Vienna, Austria). Data normalization steps such as feature transformation and standardization are needed for radiomics features due to the intrinsic differences in the range, scale, and statistical distributions of these features [
19]. All features were normalized with the Z-score method.
To reduce overfitting of the machine learning model, the least absolute shrinkage and selection operator (LASSO) algorithm with 10-fold cross validation was used to filter radiomics features [
20]. The selected features with non-zero coefficient were then linear combined that were weighted by their respective coefficients to build a radiomics signature, here we called radiomics score (Rad-score).
Establishment of machine learning models
In the training dataset, logistic regression algorithm was used to establish the machine learning model. Three models were built separately to predict LVSI in the training and testing datasets: the radiomics model, which only included the Rad-score parameter to predict LVSI; the protein model, which included COX-2 and TNC expression to predict LVSI; and the combined model, which included all selected parameters. In the testing dataset, models were tested and evaluated independently.
Statistical analysis
All statistical analyses were performed with SPSS version 25.0 (IBM Corp., Armonk, NY, USA) and R version 3.5.1. The Mann-Whitney U test is used for continuous variables, and the Pearson chi-square test is used for non-continuous variables to evaluate the feature distribution of the training dataset and the testing dataset. Spearman correlation analysis of LVSI situation with immunohistochemical data and PET imaging features was performed. Mann-Whitney U test was used to distinguish differences in PET features and protein expressions between patients with early-stage cervical cancer with (LVSI group) and without (non-LVSI group) in all dataset. Multivariate logistic regression analysis was used to predict LVSI in all patients.
Receiver operating characteristic (ROC) curve was used to evaluate the performance of each machine learning model. Pairwise comparison of area under ROC curves (AUC) was performed with DeLong test. All statistical tests were performed with a two-tailed p < 0.05 considered statistically significant.
Discussion
According to the 2018 FIGO criteria [
15], surgery is the primary treatment for patients with early-stage squamous cervical carcinoma. LVSI has an important influence on surgery and patient prognosis according to the FIGO criteria [
15]. Although most patients have excellent prognosis, approximately 30% patients might have recurrence and decreased survival rate [
21]. Many studies have proven that LVSI is closely associated with prognosis and is an independent risk factor [
2‐
4,
6,
7]. The expression of some proteins, including COX-2, TNC, and others, are related to LVSI, tumor microenvironment, and inflammation [
11,
12,
22]. Hence, accurate and early assessment of LVSI is important in prognosis assessment and treatment decision making in order to ensure that patients can obtain the maximum treatment benefit.
Our study aimed to establish a machine learning model that combines radiomics derived from PET images with molecular proteins that are associated with the pathology of cervical cancer in order to predict LVSI in patients with early-stage cervical cancer. The results indicate that the Rad-score was closely correlated with LVSI, and there were statistically significant differences in COX-2, TNC, and the Rad-score between the LVSI and non-LVSI groups. Moreover, we determined that the model based on the Rad-score could predict LVSI. When the Rad-score and molecular protein expression were combined, the AUC of the model improved a bit in the testing dataset, but the DeLong test showed no statically significant difference between the two models in the testing dataset.
Malignant tumors exhibit intratumoral biological heterogeneity and lead to changes in the texture parameters of the corresponding primary tumor on PET images. A previous study determined that heterogenic FDG uptake within a tumor correlated with intratumoral histopathological appearance [
23]. Several researches proved that the texture information of tumors reflected the tumors’ heterogeneity [
24,
25]. The selected radiomics features of our study were also texture parameters of PET image. Mu et al. [
26] found that inverse difference moment and correlation showed statistically significant differences between the early (stages I and II) and advanced stages (stages III and IV) of cervical cancer. Similarly, inverse difference moment and correlation were selected to calculate the Rad-score for predicting LVSI in our study. Recently, Li et al. [
18] study showed that the PET textures of primary tumor could predict lymphatic metastasis in early-stage cervical carcinoma (AUC = 0.757 in the validation dataset; 95% CI, 0.545–0.904;
p < 0.05). Other research have also shown that radiomics of primary tumor based on PET images could reflect tumor malignancy and were associated with nodal metastases and molecular subtypes of solid tumor [
27,
28]. The Rad-score [
29], which is calculated by the linear combination of selected features (including histogram and texture parameters) weighted by their respective coefficients selected as informative features, is usually used for radiomics analysis. In the present study, the Mann-Whitney U test showed that the Rad-score and TNC had significant differences between the LVSI and non-LVSI in all datasets, and the LVSI group had a higher Rad-score than the non-LVSI group in Fig.
5.
A previous research found that the molecular expression of some proteins was correlated with LVSI. Normal cervical tissues have weak expression of TNC and COX-2. Previously, Pilch et al. [
13] determined that, in invasive cervical carcinoma, TNC expression was markedly increased. Other studies have proven that TNC has a significant role in tumor growth, migration, metastasis, angiogenesis, and stromal inflammation [
14,
30,
31]. The study of Liu et al. [
11] found that COX-2 expression was associated with lymphangiogenesis and lymph node metastasis in cervical cancer. Similarly, Hoellen et al. [
12] proved that COX-2 expression was significantly associated with LVSI (
p = 0.017). Similarly, our study also found that the differences in COX-2 and TNC between the LVSI and non-LVSI groups were statically significant. In the multivariate logistic regression, TNC expression was associated with LVSI. Furthermore, COX-2 had a slight correlation with TNC (Fig.
4). We hypothesized that the result may be caused by the inflammatory microenvironment of the tumor. Liu et al. [
11] demonstrated that COX-2 may promote cancer progression and metastasis by enhancing the expression of vascular endothelial growth factor C and other mechanisms. The research showed that TNC could also facilitate the formation of cancer stroma, including desmoplasia and angiogenesis, and enhanced inflammation in the cancer stroma may augment macrophage recruitment and secretion of tumor-promoting and inflammatory cytokines by macrophages and fibroblasts [
14]. Thus, TNC and COX-2 were selected to predict LVSI in cervical cancer.
Although some protein expression and radiomics were associated with LVSI in cervical carcinoma, the correlation of molecular proteins and radiomics has been rarely explored and confirmed. In our study, the Rad-score was correlated with TNC, according to the Spearman correlation analysis (Fig.
4). Thus, we assumed that the PET imaging textures of primary tumors changed through TNC and other proteins. The textures of primary tumor reflect the heterogeneity of tumor, and they could be used to predict LVSI in cervical carcinoma. Another research also hoped to utilize the expression of TNC on PET imaging by devising a new PET tracer [
32]. However, our result initially showed that radiomics derived from PET imaging provided a new possibility for non-invasive visualization of TNC expression. Song et al. [
33] also found that the image signal changes on magnetic resonance imaging (MRI) were consistent with TNC expression, and cervical cancer tissues with node metastasis had the highest TNC expression.
Three machine learning models were established with logistic regression algorithm in the training dataset and evaluated in the testing dataset. All three models performed well in the training dataset (Table
4), but the radiomics model had the highest AUC in the training and external dataset (Figs.
6 and
8). However, in the testing dataset, the AUC value of the combined model was higher than that of the other models (Table
4). The reason for the results was that our dataset was slightly smaller. Thus, we used all datasets to perform the multivariate logistic regression analysis (Table
3). The results also showed that the Rad-score and TNC were associated with LVSI in all datasets. Two different methods (statistics and machine learning) both confirmed that the combination of radiomics and TNC could predict LVSI in early-stage cervical cancer. In the external dataset, the combined model for predicting LVSI was also credible and the specificity of combined model was 100%. But the AUC of radiomics model was the best among three models in the external dataset. The DeLong test also indicated that the AUC of the ROC of the combined model was better than that of the protein model in the training dataset (Table
5). Previously, a few researchers also wanted to predict or distinguish LVSI through radiology for cervical cancer. Yang et al. [
34] determined found that the minimum apparent diffusion coefficient and the minimum apparent diffusion coefficient ratio were significantly lower in LVI-positive invasive cervical cancer than in LVI-negative invasive cervical cancer (0.772 ± 0.062 vs. 0.917 ± 0.052,
p < 0.001, and 0.712 ± 0.078 × 10
− 3 vs. 0.867 ± 0.099 × 10
− 3 mm
2/s,
p < 0.001, respectively). Gross tumor volume [
35] on MRI was also identified to be a possible independent risk factor for predicting LVSI (AUC = 0.700,
p < 0.05). Recently, the use of radiomics based on magnetic resonance for predicting LVSI has been studied. According to Hua et al. [
36], the model based on multiparametric MRI showed the best prediction results, with an AUC of 0.842 (95% CI, 0.772–0.913; sensitivity = 0.773; specificity = 0.776) in the training cohort and 0.775 (95% CI, 0.637–0.912; sensitivity = 0.739; specificity = 0.667) in the validation cohort. Similarly, Li et al. [
37] also found that the radiomics nomogram derived from MRI showed favorable discrimination between LVSI and non-LVSI groups, with an AUC of 0.754 (95% CI, 0.6326–0.8745) in the training cohort and 0.727 (95% CI, 0.5449–0.9097) in the validation cohort. We initially used the combination of PET radiomics with protein molecule to predict LVSI, which showed that the radiomics and combined models based on
18F-FDG PET imaging showed better results than those of previous studies.
For patients with early-stage cervical cancer, when LVSI is evident, pelvic lymphadenectomy should be considered, along with modified radical hysterectomy according to the FIGO criteria [
15]. However, pelvic lymphadenectomy may cause a series of postoperative complications including lower limb lymphedema, bladder dysfunction and others [
38‐
40]. Combined model and radiomics model of our search could predict LVSI with PET image and histological data before surgical treatment. When the machine learning model predicts that LVSI is negative, patients are likely to avoid being subjected to pelvic lymphadenectomy. Individualized medical treatment with radiomics can improve the postoperative life quality of patients.
However, this study has some limitations. First, the size of the dataset was inadequate; thus, we need a larger number of dataset to test our models as well as multicenter imaging data to evaluate reproducibility. Second, we only analyzed the association of the expression of TNC and COX-2 with LVSI. In the future, we hope to perform more protein analyses and explore the correlation of DNA with LVSI and the function of radiogenomics in order to predict LVSI. Finally, the PET image resolution was low, thus limiting the precision of the segment of tumor VOI as well as the extraction of the radiomics features.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.