Radiomics nomogram for preoperative differentiation of early-stage serous borderline ovarian tumors and serous malignant ovarian tumors

Objectives This study aimed to construct a radiomics nomogram and validate its performance in the preoperative differentiation between early-stage (I and II) serous borderline ovarian tumors (SBOTs) and serous malignant ovarian tumors (SMOTs). Methods Data were collected from 80 patients with early-stage SBOTs and 102 with early-stage SMOTs (training set: n = 127; validation set: n = 55). Univariate and multivariate analyses were performed to identify the independent clinicoradiological factors. A radiomics signature model was constructed using radiomics features extracted from multidetector computed tomography images of the venous phase, in which the least absolute shrinkage and selection operator regression was employed to lessen the dimensionality of the data and choose the radiomics features. A nomogram model was established by combining independent clinicoradiological factors with the radiomics signature. The performance of nomogram calibration, discrimination, and clinical usefulness was evaluated using training and validation sets. Results In terms of clinicoradiological characteristics, age (p = 0.001), the diameter of the solid component (p = 0.009), and human epididymis protein 4 level (p < 0.001) were identified as the independent risk factors of SMOT, for which the area under the curves (AUCs) were calculated to be 0.850 and 0.836 in the training and validation sets, respectively. Nine features were finally selected to construct the radiomics signature model, which exhibited AUCs of 0.879 and 0.826 for the training and validation sets, respectively. The nomogram model demonstrated considerable calibration and discrimination with AUCs of 0.940 and 0.909 for the training and validation sets, respectively. The nomogram model displayed more prominent clinical usefulness than the clinicoradiological and radiomics signature models according to the decision curve analysis. Conclusions The nomogram model can be employed as an individualized preoperative non-invasive tool for differentiating early-stage SBOTs from SMOTs.


Introduction
Epithelial ovarian tumors account for 60% of all ovarian tumors and are common in women all over the world (1).Ovarian epithelial tumors can be morphologically categorized into three types as follows: benign, borderline, and malignant (2), with borderline tumors accounting for 10%-15% of all primary ovarian tumors (3,4).Typically, compared with serous malignant ovarian tumors (SMOTs), serous borderline ovarian tumors (SBOTs) have a better prognosis, implying the vitality of early diagnosis and treatment.
Over the past decade, increasing efforts have been made to differentiate between borderline ovarian tumors (BOTs) and malignant ovarian tumors (MOTs) to modify the differential diagnosis and allow for more timely and effective treatments.Nougaret et al. (5) demonstrated the efficient differentiation of SBOTs from low-grade SMOTs using multidetector computed tomography (MDCT) features, such as peritoneal disease patterns and solid tumor volumes.Furthermore, Grabowska-Derlatka et al. (6) indicated that the morphological assessment of tumor vascularity in MDCT is an efficient approach to distinguish BOTs from MOTs.Previous studies demonstrated that BOTs can be distinguished from MOTs by analyzing the apparent diffusion coefficient of a solid tumor component using magnetic resonance imaging (MRI) (7,8).Differentiating BOTs from MOTs based on imaging alone is generally difficult; thus, some studies have attempted to predict BOTs using preoperative imaging and tumor biomarkers.A previous study validated the superior performance of combining MRI with serum cancer antigen 125 (CA125) for differentiating between BOTs and MOTs (9).In our previous study, age, vascular abnormalities, and largest solid portion were independent factors distinguishing SBOTs from SMOTs (10).
Radiomics is an emerging translational research field that aims to extract high-dimensional quantitative features from clinical images and provide useful information (11).This approach has been adopted in a variety of applications, including the prediction of treatment response and outcome, tumor staging, tissue identification, and cancer genetics assessment (12, 13).SBOTs and SMOTs share similar clinical symptoms and imaging features in their early stages, making differentiation between these two entities difficult.We developed a nomogram combining clinical characteristics, conventional imaging characteristics, and MDCT radiomics features for preoperative differentiation between earlystage SBOTs and SMOTs.As far as we know, it is the first nomogram model established for preoperative differentiation between early-stage SBOTs and SMOTs.

Patients
This retrospective study was conducted and approved by the institutional review committee, with the requirement of informed consent waived.Eighty SBOT patients (54 stage I and 26 stage II) and 102 SMOT patients (63 stage I and 39 stage II) from December 2017 to June 2020 were enrolled from our hospital information system (HIS) database into this analysis.The inclusion criteria were as follows: 1) histopathologically confirmed SBOTs and SMOTs after surgery and 2) SBOTs and SMOTs diagnosed as stage I or II based on the International Federation of Gynecology and Obstetrics (FIGO) guideline.The exclusion criteria were 1) SBOTs and SMOTs of stage III or IV according to the FIGO guideline, 2) patients that have accepted preoperative therapy (radiotherapy, chemotherapy, or chemoradiotherapy) before MDCT examination, and 3) incomplete clinical data or poor imaging data.The cohort was divided into a training set (127 patients) and a validation set (55 patients) according to the examination date of MDCT, with the total process depicted in Figure 1.

Clinicoradiological risk factors
The clinical characteristics of age, CA125 level (≤50 U/mL; >50 U/mL) (9), and human epididymis protein 4 (HE4) level (≤150 pmol/L; >150 pmol/L) (14) derived from medical records were recorded in the HIS.The imaging was viewed separately and independently by two radiologists in abdominal diagnosis, with 8 and 9 years of experience who were blinded to the clinical variables and pathological diagnosis.The following data of conventional imaging characteristics were recorded: the diameter of the tumor on the axial MDCT image, the diameter of the solid component on the axial MDCT image, location (unilateral or bilateral), texture (predominantly cystic or solid), margins (smooth or irregular), ascites (present or absent), and vascular abnormality (present or absent; vascular abnormality was defined as the presence of no less than one of the following situations: a) chaotic or serpentine course of the vessel, b) microaneurysm, and c) arteriovenous fistula).

MDCT radiomics analysis
The radiomics analysis methodology was carried out as described in a previous paper (15).In brief, regions of interest (ROIs) were manually segmented with 3D slicer, a free open-source software (https://www.slicer.org/).Afterward, 7 categories and 1,167 radiomics features were extracted with the "radiomics" package of the 3D slicer.
All radiomics features extracted from MDCT of VP were normalized.Significant radiomics features in the training set were selected applying interclass correlation coefficient and least absolute shrinkage and selection operator (LASSO) regression.A radiomics score (Rad-score) of each patient was calculated by means of a linear combination of the chosen features weighted by their coefficients.

Statistical analysis
All statistical analyses were carried out on SPSS (version 20, Chicago, IL, USA) and R (https://www.r-project.org/).Statistical significance was defined as p <0.05.

Patient characteristics
The clinicoradiological characteristics of all patients in the training and validation sets are provided in Table 1.SBOTs and SMOTs exhibited significant differences in age, HE4 level, location, texture, margins, vascular abnormalities, and the diameter of the solid component (p < 0.05), while there were no significant differences in CA125 level, ascites, and size (p > 0.05) in the training set.Age (p = 0.001), the diameter of the solid component (p = 0.009), and HE4 level (p < 0.001) were identified as the independent factors in the clinicoradiological model according to multiple logistic regression analysis.Tumors with older age [odds ratio (OR), 1.11; 95% CI, 1.05-1.18],larger diameter of the solid component (OR, 1.07; 95% CI, 1.02-1.13),and higher HE4 level (OR, 56.94; 95% CI, 7.1-456.49)trended to develop to SMOTs.The flowchart of the selection of patients.

Clinical use of the nomogram model
Age, the diameter of the solid component, HE4 level, and Radscore were integrated for the nomogram model construction (Figure 3).The calibration curve of the nomogram in Figure 4 demonstrated considerable calibration in both the training set (p = 0.994) and the validation set (p = 0.859).The performance of the clinicoradiological model, the radiomics signature model, and the

Discussion
The biological characteristics, prognoses, and treatments completely vary between SBOTs and SMOTs, implying the crucial role of accurate preoperative differentiation in determining individualized treatment options and improving postoperative quality of life (2).FIGO stage III or IV ovarian tumors more prominently display aggressive characteristics, such as lymphatic and distant metastases, resulting in a tendency to be identified as SMOT (16,17).Therefore, patients in the early stages were employed as subjects in our study, which holds greater significance in clinical practice.
Age is generally acknowledged as one of the factors that discriminate between SBOTs and SMOTs (18).As indicated by our previous study, the median age of the patients with SMOTs was 10 years older than patients with SBOTs, showing a statistically significant difference (10).In the present study, using multiple logistic regression analysis, age was validated as an independent predictor in accordance with previous studies (18).
In the past two decades, studies on several serum biomarkers for the diagnosis of epithelial ovarian cancer including CA125, carbohydrate antigen 19-9, carcinoembryonic antigen, and HE4 have already been conducted (19, 20).HE4 is a relatively reliable biomarker for ovarian cancer detection, whereas CA125 is commonly used for early ovarian cancer diagnosis in clinical practice (14).In terms of ovarian cancer detection, HE4 exhibits superior diagnostic performance to CA125, while their combination provides higher sensitivity at the expense of lower specificity compared with HE4 alone (21,22).In the present study, the HE4 level was proven to be an independent predictor in the clinicoradiological model, showing a superior diagnostic performance to the CA125 level in terms of differentiation between early-stage SBOTs and SMOTs.
In routine clinical practice, MDCT is the most commonly used imaging method for preoperative assessment and postoperative surveillance of patients with ovarian tumors.Previous studies have indicated that MDCT and MRI are promising imaging tools for differentiating BOTs from MOTs (23,24).Due to the focus on variables such as tumor shape, tumor volume, and tumor vascular changes in early studies (6,17), it is difficult to distinguish BOTs from MOTs depending on imaging information alone, despite the medical imaging methods used to distinguish them, which could result from some morphological imaging findings overlapping between BOTs and MOTs, such as irregularly thickened walls or septa, predominantly solid masses, and vascular abnormalities.Therefore, medical imaging methods must be combined with clinical data to achieve a differential diagnosis.Our previous study demonstrated the validity of the combined analysis of age, vascular abnormalities, and the diameter of the solid components for differentiating between SBOTs and SMOTs (10).Shin et al. ( 25) found that a more reproducible and accurate differentiation of BOTs from MOTs could be achieved by combining MDCT data and CA125 levels.
With advancements in computer technology, radiomics has emerged as a new research method in recent years.Medical images can be converted into high-dimensional mineable data after the extraction of quantitative image features.Radiomics could contribute to exploring the potential connections related to tumor occurrence and development offering reference for clinicians to make a proper diagnosis before surgery.This is a promising field used in oncology and has been adopted in the differential diagnosis of pulmonary solid nodule (26), renal masses (27), liver tumor (28), and thyroid follicular neoplasms (29) on MDCT images.
In our previous study (15), the MDCT-based radiomics model was efficient in discriminating SBOTs from SMOTs, with AUCs for the radiomics models AP, VP, and EP of 0.8, 0.86, and 0.73, respectively; the AUC was higher for VP.In the present study, we investigated the performance of a nomogram model based on radiomics of the VP for preoperative differentiation between SBOTs and SMOTs.This study had some limitations.First, it was a single-center scanner study; a multicenter study is warranted to analyze and verify our results in the future.Second, the late stages of SBOMs and STOMs were not included, which may limit the model generalization.Third, there may be an inevitable selection bias in retrospective studies, and prospective and external validation studies are needed.Fourth, the lack of multiple segmentations affected the quality of the study, limiting the reproducibility of the results.

Conclusions
In summary, a nomogram model combining clinicoradiological risk factors and radiomics features was established in the present study and can be adopted as an individualized preoperative noninvasive tool for distinguishing between early-stage SBOTs and SMOTs.
Three models were constructed by multivariate logistic regression analysis: the clinicoradiological model by prominent clinical characteristics and conventional imaging characteristics, the radiomics signature model by the selected radiomics features, and the nomogram model by the independent clinicoradiological risk factors combined with the radiomics signature.Evaluation and validation of the different modelsThe evaluation of calibration of the nomogram model was conducted in the training and validation sets using the Hosmer-Lemeshow goodness-of-fit test.The performance of the three models in diagnosis was evaluated in the training set and validated in the validation set by plotting the receiver operating characteristic (ROC) curve and calculating the area under the ROC curve (AUC) and the multiple comparison of AUC curves using the DeLong test on Bonferroni-adjusted p-values.The AUC with 95% confidence interval (CI), sensitivity, and specificity was calculated.Decision curve analysis (DCA) was performed according to net benefits and the corresponding threshold probabilities in the training and validation sets, to assess the clinical application of the nomogram model.Our study workflow diagram is shown in Figure2.
nomogram model in diagnosis is evaluated and summarized in Table 2, and the ROC curves are depicted in Figure 5.The nomogram model exhibited significantly higher AUC compared with the clinicoradiological model and the radiomics signature model (p < 0.001) in the training set, while there was no significant difference in AUC between the nomogram model and the radiomics signature (p = 0.123) in the validation set.In the training (p = 0.494) and validation sets (p = 0.8972), no significant differences in AUC were revealed between the clinicoradiological model and the radiomics signature model.The DCA of the three models (Figure 6) demonstrated a more prominent overall net benefit of the nomogram model in differentiating between SBOTs and SMOTs in comparison to the clinicoradiological model and the radiomics signature model.

FIGURE 3 FIGURE 4
FIGURE 3The radiomics nomogram, combining age, the diameter of the solid component, HE4 level, and Rad-score, developed in the training set.

5 ROC
FIGURE 5 ROC curve of the clinicoradiological model, the radiomics signature model, and the nomogram model in the training (A) and validation (B) sets.

TABLE 1
Comparison of the clinicoradiological characteristics of SBOT and SMOT patients.

TABLE 2
Predictive performance of the three models.