Diagnostic value of a CT-based radiomics nomogram for discrimination of benign and early stage malignant ovarian tumors

Ovarian cancer is one of the three common malignant tumors of the female reproductive system. In 2018, there were about 295,414 new cases and 184,799 ovarian cancer-related deaths worldwide, while only about 25% of ovarian cancer patients were diagnosed at early stage [1]. Imaging examination is vital for tumor detection, localization, benign and malignant determination, staging evaluation of malignant tumors, and prognosis of patients with ovarian cancer. However, early stage ovarian malignancies (FIGO stage I/II) are relatively insidious and has no obvious clinical symptoms. In clinical practice, the experience and professional level of imaging departments and clinicians noticeably influence the early diagnosis of ovarian cancer. In particular, the imaging features of benign and early stage malignant ovarian tumors often overlap. To identify only basing on the clinical manifestations, tumor markers, and the traditional computed tomography (CT) manifestations recognized by the naked eye, is lacking repeatability and objectivity, and often fails to accurately identify the nature of the lesion.

Although pathological diagnosis is the gold standard for distinguishing benign and malignant tumors, it belongs to an invasive examination. Sampling can only reflect the local condition rather than the entire tumor, and there are limitations in operation and the risk of tumor dissemination. In China, as a routine diagnostic method for ovarian cancer, CT possesses the advantages of being non-invasive, fast, and better presentation of morphological characteristics of ovarian lesions. It has markedly contributed to the diagnosis, as well as determination of the treatment protocols and evaluation of therapeutic effects [2]. Radiomics has noticeably attracted radiologists’ attention in recent years. The tumor heterogeneity is assessed objectively and quantitatively by radiomics through extracting the developable high-dimensional imaging features (e.g., intensity, geometry, texture, etc.) from medical images, and then, using a series of statistical tools and algorithms to quantitatively analyze the extracted features [3, 4]. To date, radiomics has been widely used in the study of diverse types of cancer. Wei et al. [5] constructed a new radiomics model using imaging phenotype and clinical variables to predict the overall survival time (OS) of hepatocellular carcinoma (HCC) patients receiving stereotactic body radiation therapy (SBRT). Huang et al. [6] developed an ultrasound-based radiomics model to distinguish between sclerosing adenopathy (SA) and invasive ductal carcinoma (IDC) to avoid misdiagnosis and unnecessary biopsy. Ramtohul et al. [7] evaluated whether the radiological characteristics based on multi parameter dynamic enhanced MRI could help to distinguish the expression of HER2 in breast cancer. The results showed that the radiological characteristics and tumor descriptors from multi parameter dynamic enhanced MRI could predict different HER2 expressions in breast cancer with therapeutic significance. Kang et al. [8] demonstrated that a radiomics nomogram model based on CT images can predict the recurrence and metastasis of clear cell renal cell carcinoma.

In recent years, many studies have shown that radiomics nomograms can provide valuable and reliable information for ovarian tumors. The radiomic nomogram was first used in a 2016 study published in the Journal of Clinical oncology. The study combined selected radiomic features with clinically relevant risk factors to construct nomograms. This prediction model can effectively conduct preoperative personalized prediction for patients with colorectal cancer lymph node metastasis and build a stable and feasible prediction model [9]. Radiomic nomogram analysis can be performed on medical images from different modes, such as CT, MRI, and ultrasound. Pan et al. [10] constructed a combined nomogram model based on radiological and semantic features for preoperative classification of serous and mucinous pathological types in ovarian cystadenoma patients. Hu et al. [11] investigated a nomogram based on arterial phase CT radiomics features and clinical features to help distinguish primary and secondary ovarian cancer. Li et al. [12] explored the difference in the application potential of two-dimensional (2D) and three-dimensional (3D) radiomics models based on non-contrast CT scans in differentiating benign and malignant ovarian tumors. The results showed that 2D and 3D radiomics nomogram models had comparable diagnostic efficacy in the differential diagnosis of benign and malignant ovarian tumors. Li et al. [13] combined radiomic characteristics and clinical factors to construct a radiomic nomogram based on venous phase CT, which has good clinical application value in the differential diagnosis of ovarian cystadenoma and endometriosis cyst. Zhang et al. [14] constructed a nomogram model combining radscore and clinical features, which can be used to detect synchronous ovarian metastasis in female gastric cancer patients. To differentiate benign, borderline, and malignant ovarian serous tumors, Qi et al. [15] integrated novel radiomics signatures from ultrasound and clinical factors to create a nomogram, thereby reducing or avoiding the risk of biopsy and surgery. Yao et al. [16] constructed a clinical–radiomics nomogram, which applied ultrasound radiomics features to distinguish the histopathological types of EOC for the first time, helping gynecologists to identify the types of EOC noninvasively before surgery. Xu et al. [17] used the image nomogram based on diffusion weighted imaging (DWI) and apparent diffusion coefficient (ADC) to classify epithelial ovarian tumors, which has significant clinical significance. CT is the most important imaging method for assessing the extent of ovarian tumors, facilitating preoperative tumor staging and rational surgical planning [18]. However, there are few studies on CT-based radiomics in early ovarian malignancies. Therefore, the present study aimed to explore the diagnostic value of a CT-based radiomics nomogram for discrimination of benign and early stage malignant ovarian tumors.

The prognosis of ovarian malignant tumors depends on early stage diagnosis, surgical treatment, and postoperative systemic treatment. Therefore, the early and accurate identification of benign, malignant, and aggressive lesions is vital for the selection of an appropriate treatment option and prediction of prognosis of patients with ovarian cancer. However, the current routine gynecological examinations, traditional imaging features, and tumor markers all have certain difficulties in the qualitative diagnosis of early stage ovarian tumors. Therefore, how to use non-invasive methods to improve the diagnostic accuracy of early stage ovarian cancer has always been a hot topic for gynecologists. In this study, we provided three diagnostic models: a radiomics model constructed with optimal features, a traditional model that combines clinical manifestations, tumor markers, and traditional CT manifestations recognized by the naked eye, and a nomogram model that combines important traditional factors and radiomics features.

Radiomics extracts deep information that cannot be recognized by the human eye through in-depth exploration of high-dimensional features of CT images, and then, quantitatively analyzes the tumor heterogeneity, which can reflect tumor information more objectively and comprehensively [19]. In recent years, radiomics has been used to non-invasively identify benign and malignant ovarian tumors [20, 21], as well as for the purposes of histological grading [22], evaluation of molecular typing [23, 24], assessment of efficacy [25], and prediction of metastasis [26] and prognosis [27, 28]. At present, there is no relevant research on the use of CT imaging to distinguish benign and early stage malignant ovarian tumors, and research have mainly concentrated on the contrast-enhanced CT or MRI.

P-CT refers to scanning without intravenous injection of iodine containing contrast agent, while CE-CT refers to scanning under intravenous injection of iodine containing contrast agent. P-CT can provide basic anatomical structure information of the examination area, but usually cannot distinguish between benign and malignant lesions. CE-CT can clearly display the location, shape, range, internal components, blood supply, and presence or absence of metastasis of ovarian tumors. It is the preferred imaging examination method for preoperative FIGO staging and treatment planning of ovarian tumors. P-CT and CE-CT can provide different information, but both are indispensable components of ovarian tumors CT examination. Before performing enhanced scanning, P-CT scan is necessary. The differentiation of ovarian tumor tissues is associated with the characteristics of gray value. The radiomics of the P-CT scan has a potential value in the classification of ovarian tumor tissues. Therefore, utilization of the P-CT scan for tumor identification is logical in theory, and its working principle is easier than the CE-CT [22]. Previous studies [22, 26] have also shown that the radiomics model based on the P-CT images has a high diagnostic efficiency, which could be significant for the identification and prediction of ovarian benign and malignant tumors. Compared with P-CT scans, CE-CT scans can provide more valuable data and more comprehensively reflect the heterogeneity of tumors; nevertheless, the CE-CT is susceptible to the subjective influences of the contrast agent itself and an operator’s experience, and the results may be biased [29]. A previous research [26] showed that a nomogram based on venous CT radiomic has a promising efficacy in predicting lymph node metastasis in high-grade serous carcinoma. Hence, in the present study, we used the P-CT and CE-CT to identify further accurately benign and early stage malignant ovarian tumors, so that patients can receive earlier and more personalized treatment options without increasing the economic burden. In this study, a total of 604 quantitative texture features were extracted from P-CT and CE-CT images. Table 2 provides the diagnostic performance of the model constructed based on these radiomics features. The results showed that the AUC values of 1.00–0.90 for the combined diagnosis were evaluated as excellent [30], but the features were too redundant. Therefore, this article reduces the dimensionality of the combined features and ultimately obtains 6 optimal features, including 2 features from P-CT and 4 features from CE-CT, for Radscore calculation to identify benign and early malignant ovarian tumors.

In our study, a total of 6 radiomics features were obtained using the combined scanning, including S (0, 1) Correlat, S (3, 3) Correlat, Perc. 10%, Perc. 90%, S (0, 2) SumEntrp, and WavEnLL_s-4. S (0,1) Correlat and S (3,3) Correlat were correlated together, belonging to the characteristics of the GLCM. The correlation describes the degree of similarity of the GLCM in the row and column directions. When the correlation value is relatively uniform, the degree of similarity in the rows and columns is large [31]. S (0,2) SumEntrp indicates the sum of entropy, which is also a feature of the GLCM, and it could be used to describe the degree of textural complexity. The larger the entropy value, the more uneven the texture of the studied image [32]. Perc.10% and Perc.90% belong to the histogram feature parameters, and the histogram parameters are based on the distribution of the voxel intensity of the pixel distribution in the CT image to reflect the tissue heterogeneity. The larger the value, the more obvious the tumor heterogeneity [33]. WavEnLL_s-4 is the low-frequency wavelet coefficient, which is the low-frequency component of the wavelet transformation model, highlighting the low-spatial frequency component, and it is a high-order feature. Wavelet transform is an image processing technique using a combination of high-frequency and low-frequency bandpass filters to decompose an image to obtain important information that may be hidden from the image [34]. At present, high-order statistics are rarely used in radiomics, and the clinical significance of the parameters still needs to be further explored. The histogram of the Radscore values of the 6 optimal radiomics features shows that the Radscore values of the benign group are lower than those of the early malignant group, indicating that the tumor heterogeneity is more significant in the early malignant group. Table 3 shows that the radiomics model constructed using the 6 optimal radiomics features has high AUC values, sensitivity, specificity, and accuracy, indicating its very good diagnostic performance (AUC0.90–0.80) [30]; the AUC value in the training set (0.856) is slightly higher than the value in the validation data set (0.843), indicating that the radiomics model has strong generalization ability.

Univariate analysis showed that among the traditional diagnostic factors, Ca125, HE4, and the largest tumor diameter were correlated with benign and early malignant ovarian tumors (P < 0.1). Multivariate logistic regression analysis revealed that imaging features have a certain diagnostic value for benign and early stage malignant ovarian tumors, while they cannot be independent predictors. It was speculated that CT imaging features overlap between benign and early stage malignant ovarian tumors. The independent predictors of early stage malignant tumors were Ca125 and HE4 (P < 0.05), which was consistent with the results of higher preoperative levels of Ca125 and HE4 in patients with ovarian malignant tumors in a previous study [35]. Therefore, based on Ca125 and HE4, the present study constructed a traditional diagnostic model for ovarian benign and early stage malignant tumors. Compared with traditional diagnostic models, the radiomics model in the validation set increased AUC by 6.3% (AUC, 0.843 vs. 0.780, P = 0.435), but the difference was not significant. Compared with traditional models, the radiomics model has higher sensitivity (0.955 vs. 0.731) and lower specificity (0.667 vs. 0.810), indicating fewer missed diagnoses and ensuring diagnostic quality. However, the degree of intention was relatively low, and it was easy to be misdiagnosed; if these two methods are combined, they can compensate for their respective shortcomings and make the diagnosis to be more accurate.

Furthermore, we combined the traditional diagnostic model (Ca125 and HE4) and the radiomics model to construct a nomogram, which easy-to-use and showed an excellent predictive performance in both the training and the validation data sets. In the validation data set, using the traditional diagnostic model, the radiomics model, and the nomogram, the AUC values for predicting benign and early stage malignant ovarian tumors were 0.780, 0.843, and 0.913, respectively, which were all greater than 0.7, demonstrating that the models have a good diagnostic value [30], and the nomogram has the highest predictive capability. Compared with traditional diagnostic models and radiomics models, the nomogram showed an increase in AUC of 10.3% (0.910 vs. 0.807, P = 0.000) and 5.4% (0.910 vs. 0.856, P = 0.012) in the training data set, respectively; in the validation data set, AUC increased by 13.3% (0.913 vs. 0.780, P = 0.003) and 7.0% (0.843 vs. 0.913, P = 0.108), respectively. This indicates that compared to radiomics models, the nomogram has higher performance in the training data set, but weaker generalization ability, and there is no statistically significant difference in the validation data set; compared with traditional diagnostic models, nomograms exhibit higher performance in both training and validation data sets, and the differences are statistically significant. The robustness is verified through validation data sets. To evaluate the clinical usefulness of the radiomics nomogram, decision curve analysis (DCA) was applied in this study, which is a novel method to calculate the net benefit at various threshold probabilities to insight clinical consequences. DCA revealed that the radiomics nomogram has a greater clinical value in the discrimination of benign and early stage malignant ovarian tumors within the threshold range of 0.4–1.0, confirming a promising clinical utility. Our study found that the radiomics nomogram model outperformed in diagnostic accuracy, which consistent with a recently published study that arterial phase CT imaging feature and clinical feature to distinguish primary and secondary ovarian cancer. The results of that research showed that the combination of clinical factors and arterial phase CT radiomics features was more efficient than using them alone [11]. Another study used radiomics to identify benign and malignant bone tumors, which showed that the radiomics nomogram model included clinical and radiomics features performed well in both training and validation data sets. The AUC, DCA, and net reclassification improvement (NRI) revealed that compared with the clinical model, the radiomics nomogram model exhibited a better diagnostic performance, and it has a greater clinical net benefit than the pure clinical and radiomics model [36].

The limitations of this study should be pointed out. First, the sample size was small, the pathological type of distribution was unbalanced, and the results need to be independently verified by a larger sample size and involvement of multiple centers. Second, the thickness of the scan was 5 mm, which was theoretically correct for the second order and high level of the extracted lesions. The value of first-order image features might have an impact. Finally, this was a retrospective study, and there was an inevitable selection bias.

In summary, the radiomics model could discriminate benign and early stage malignant ovarian tumors, and the radiomics nomogram model that combines traditional and radiomics models shows the best diagnostic performance, which is valuable for developing personalized treatment plans.