A preoperative nomogram incorporating CT to predict the probability of ovarian clear cell carcinoma

The final study population consisted of 533 patients (median age, 57 years; range, 44 – 70), as follows: 61/533 (11.45%) with OCCC, 354/533 (66.41%) with HGSC, 37/533 (6.94%) with endometroid OC, 17/533 (3.18%) with mucinous OC, and 64/533 (12.00%) with other OC. The clinical and laboratory data of patients with OCCC and non-OCCC are summarized and compared in Table 1 . A history of endometriosis was found significantly more often in patients with OCCC (12/61 [20%] vs 17/472 [3.6%], P < 0.001). CA-125 levels were significantly higher in patients with non-OCCC (351 [110, 1,247] vs 70 [27,176], P < 0.001). No significant difference in patient age, ethnicity, menopausal status, history of adenomyosis, or CEA level was found between patients with OCCC and those with non-OCCC. Although it was not statistically significant, there was a difference in the ethnicity of patients with and without OCCC; 20% of the OCCC population were Asian, whereas only 8% of the non-OCCC population were Asian. The imaging features identified for OCCC and non-OCCC by reader 1 and reader 2, along with kappa statistics, are shown in Table 2 . For both readers, significant differences in the prevalence of some imaging features were found between OCCC and non-OCCC. Unilateral lesions were detected significantly more often in patients with OCCC (P < 0.001). In addition, in patients with OCCC, both the mean largest ovarian tumor diameter and the mean largest cystic component diameter were significantly greater (P < 0.001 for both comparisons; see Table 2 for additional details). In patients with non-OCCC, the following imaging features were significantly more common: ascites, peritoneal implants including omental cake ( Table 2 ). A frozen pelvis was rarely found by reader 1, and its prevalence on imaging was not significantly different between patients with OCCC and those with non-OCCC ( Table 2 ). Inter-reader agreement was almost perfect (κ = 0.87) in the detection of omental cake, while it was substantial in the detection of any ascites and peritoneal implants and in the characterization of ovarian lesion morphology (κ = 0.65, κ = 0.81, κ = 0.72, respectively). Inter-reader agreement was moderate in the characterization of laterality, lymphadenopathy, calcification and 5-point subjective assessment (κ =0.62, κ = 0.66, κ = 0.35 and κ = 0.580, respectively). It was slight or fair for the detection of the remaining features ( Table 2 ). The following variables were incorporated in the most comprehensive nomogram model: age, history of endometriosis, CA-125, laterality (uni- or bilateral), peritoneal implants, morphology of the ovarian lesion, diameter of the lesion and of the largest solid component ( Figure 2 ). The diameter of the lesion was modeled as a restricted cubic spline and shows a clear nonlinear effect. The AUC of the model was 0.9 (95% CI: 0.847, 0.949) with an IPA of 33.8%. The calibration curve appears in Figure 3 . Additionally, we evaluated the performance of radiologists based on their subjective impressions provided using an ordinal 5-point scale of predicted risk. The AUC of reader 1 was 0.8 (95% CI: 0.822, 0.891) with an IPA of 22%, and the AUC of reader 2 was 0.9 (95% CI: 0.869, 0.936) with an IPA of 40%.

The institutional review board at Memorial Sloan Kettering Cancer Center approved this retrospective, Health Insurance Portability and Accountability Act-compliant study and waived the informed consent requirement. We searched our gynecologic surgery and radiology databases for patients who underwent primary debulking surgery for ovarian cancer from 1/1/2000 to 12/31/2016 with CECT of the abdomen and pelvis performed up to 90 days before primary surgery, acquired either at our institution or externally but reviewed by a radiologist at our institution. The exclusion criteria were (a) CECT obtained after neoadjuvant therapy and (b) incomplete CECT or poor CECT image quality, including scans on film. The patient selection process is summarized in Figure 1 . The clinical data of patients were obtained from a detailed review of the electronic medical record using a standardized form. The following clinical data were reviewed: age at the time of diagnosis, ethnicity, menopausal status, and clinical history of endometriosis or adenomyosis. Regarding laboratory data, the levels of cancer antigen 125 (CA-125) and carcinoembryonic antigen (CEA) at the time of diagnosis were reviewed. Two board-certificated radiologists with 5 (PCA, reader 1) and 8 (NH, reader 2) years of experience in oncologic body imaging independently analyzed the scans. The readers were blinded to clinical data, previous imaging reports, laboratory results, and histopathological results. The cases were randomly ordered, and the radiologists reviewed the CECTs on a GE Centricity Picture Archiving and Communication System (General Electric, Milwaukee, WI). Of note, CTs were performed at our and other institutions using different scanners and protocols over a long period of time. Still, we only included those available with contrast-enhanced portal-venous phase and oral contrast. The following radiological features were evaluated: Ascites, classified as small, moderate, or large. Adhesions between pelvic organs (also referred to as pelvic adhesions), characterized by bowel tethering and/or abnormal angulation of uterus and/or pelvic organs, defining a frozen pelvis. Peritoneal implants, classified according to location (upper abdomen until the iliac crests, pelvis or both), including omental cake. Lymphadenopathy, classified according to location (supra-diaphragmatic, retroperitoneal infrarenal, retroperitoneal suprarenal, intraperitoneal, pelvic or inguinal). The criterion used to identify a lymph node as suspicious for malignancy was a size of 10 mm or larger in the short axis [ 30 ]; this applied to all nodal chains, except for supra-diaphragmatic and inguinal nodes, for which the cutoffs were 5 mm and 15 mm [ 31 ], respectively. Laterality of ovarian lesion. Largest tumor diameter, as well as diameters of the largest solid and cystic components. Morphology of ovarian lesion, classified as predominantly solid, predominantly cystic unilocular or predominantly cystic multilocular [ 32 ]. Margin of lesion: smooth or irregular. Presence of calcification. At the end of the review, each reader assigned a subjective impression of the diagnosis using a five-point scale (1, definitely OCCC; 2, probably OCCC; 3, indeterminate; 4, probably non-OCCC; and 5, definitely non-OCCC) based on their expertise and clinical judgment. In patients with bilateral lesions, only the imaging features of the largest lesion were included in the analysis. The reference standard was the histopathological report based on the surgical specimens. All histopathological analyses were done by specialized gynecological pathologists in a standardized manner. Comparisons of clinical, laboratory and radiological features (within each reader) between OCCC and non-OCCC were made using Pearson's chi-squared test for categorical variables and the Mann-Whitney U-test for continuous variables. Agreement between the two radiologists was analyzed for each radiological feature by using unweighted kappa values. The guidelines of Landis and Koch [ 33 ] were followed for the interpretation of these values (0.00–0.20, slight agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, substantial agreement; 0.81–1.00, almost perfect agreement). A multivariable logistic regression model was created using selected clinical and laboratory variables and radiological variables (based on the findings of the radiologists - the individual reader datasets were stacked together to form a large dynamic dataset). One model was created using clinical, laboratory, and radiology variables together. Restricted cubic splines were used for continuous variables to relax linearity assumptions, and natural logarithm transformations were used for skewed continuous variables, when appropriate. The statistical prediction model was measured by the area under the curve (AUC) and the index of prediction accuracy (IPA) [ 34 ]. The model was internally validated with 1,000 bootstrap samples to obtain bias-corrected discrimination using a concordance index (c-index). The radiologist and model performance levels were compared; specifically, this was done by comparing the model's prediction with that of each radiologist's subjective impression, where the subjective impression was treated as an ordinal 5-point scale of predicted risk of OCCC, calculated using bootstrapping with 1000 repetitions. Calibration of the model was assessed by a visual inspection of the plots of predicted OCCC versus actual outcome. All statistical analyses were performed using the statistical program R 3.3.2 (R Foundation for Statistical Computing, Vienna, Austria). The level of statistical significance was set at p < 0.05.

In our study population, several clinical and laboratory features and a number of radiological features differed significantly between patients with OCCC and those with non-OCCC. Most of these radiological features were identified with moderate inter-reader agreement. Among clinical and laboratory features, a history of endometriosis was significantly more common in patients with OCCC, while CA-125 levels were significantly higher in patients with non-OCCC. Four imaging features were significantly more likely to be associated with OCCC; specifically, OCCC were more likely to be unilateral and to have smooth margins, were larger and were more likely to be predominantly cystic. Based on our results, we developed a preoperative nomogram for predicting the diagnosis of OCCC versus non-OCCC that combined seven readily accessible clinical, laboratory, and CECT imaging variables: age, history of endometriosis, CA-125, presence of peritoneal implants, the morphology and the diameter of the largest ovarian lesion, and the largest diameter of the solid component. This nomogram performed comparably to the subjective impressions of the experienced radiologists. Overall, inter-reader agreement for the radiological features used in the nomogram was reasonably good, being substantial for both peritoneal implants and lesion morphology and moderate for lymphadenopathy; however, the inter-reader agreement for lesion diameter was slight. To the best of our knowledge, ours is the first study to generate a preoperative nomogram to predict the probability of OCCC versus non-OCCC based on clinical, laboratory, and radiological features on CECT. The clinical impact of our study lies in building the first preoperative nomogram for predicting the diagnosis of OCCC, which could be used to help guide the treatment approach. Few studies have compared clinical, laboratory, and radiological features between patients with OCCC and those with non-OCCC. With regard to clinical features, our results are in line with those of previous reports, which did not show significant differences in patient age or menopausal status between these two groups of patients [ 35 ; 36 ]. However, in a study with 28,082 patients with epithelial OC, including 1,411 with OCCC, the median age at diagnosis was significantly lower for those patients with OCCC [ 37 ; 38 ]. Similarly to our study, prior studies found that significantly higher proportions of patients with OCCC had a history of endometriosis [ 37 ; 39 ]. Unlike our study, a study by Tanaka et al. [ 36 ] found no significant differences in tumor markers. Regarding imaging features, other investigators have also found OCCC tumors to be more frequently unilateral and smooth, larger, with bigger cystic components [ 17 - 22 ], and less often associated with peritoneal disease when compared with non-OCCC [ 35 ; 36 ]. Indeed, in an analysis of 48 patients with OCCC evaluated with either CT or MRI, Joo et al. found that most patients had a history of endometriosis and had unilateral disease; furthermore, most had predominantly cystic masses, unilocular in architecture, with smooth margins [ 40 ]. In our study, 26% to 48% of patients with OCCC had lymphadenopathy identified on CECT, which is comparable with a rate of nodal metastasis of 24.4% found in another study based on lymph node dissection [ 9 ]. On the other hand, Ma et al. [ 35 ] did not find any significant difference in lymphadenopathy between patients with OCCC and those with HGSC; however, the imaging modality used in their study was pelvic MRI, and not all lymphatic chains were included. Our study had some limitations. We performed bootstrap validation to avoid over-interpretation of the results, but we did not apply external validation. Thus, the generalizability of our results is limited. Second, we only analyzed cases of ovarian cancer; we did not have a non-cancer control group of patients with benign ovarian cysts, pelvic endometriosis, and other benign pelvic pathologies. Finally, since the nomogram was not applied prospectively, outcome date were not collected, including surgical morbidity, complications, days of hospitalization and survival data. Consequently, additional studies (particularly prospective studies) are needed to overcome these limitations and further evaluate the clinical value of the nomogram among different patient populations. In conclusion, our study demonstrated that a presurgical nomogram model with readily accessible clinical, laboratory, and CECT imaging variables was a powerful tool to predict the diagnosis of OCCC. Thus, along with having established roles in staging, preoperative CECT may have a role in prediction of OCCC; such prediction could potentially be used to aid in treatment planning.

Ovarian cancer (OC) accounts for approximately 5% of cancer deaths among women in the United States, causing more deaths than any other gynecologic cancer [ 1 ]. The American Cancer Society estimates that approximately 13,270 women in the United States will die from ovarian cancer in 2023 [ 2 ]. The World Health Organization currently distinguishes five principal histologic subtypes of ovarian cancer: high-grade serous carcinoma (HGSC) and low-grade serous, endometrioid, clear cell, and mucinous carcinomas [ 3 ]. Of these various subtypes, HGSC is the most common, accounting for the vast majority of newly diagnosed OC. Ovarian clear cell carcinoma (OCCC) is the most common subtype after HGSC, accounting for 5%-25% of OC depending on the population of interest. Compared to other subtypes, OCCC is more frequently diagnosed among Asian women [ 4 ] and at a younger age, and is more often associated with a history of endometriosis [ 5 ; 6 ]. Patients with OCCC are often diagnosed earlier compared to those with HGSC, which usually manifests at an advanced stage. If diagnosed early, patients with OCCC have a good prognosis, with a 3-year overall survival rate of 76%-100% [ 7 ; 8 ] and a 5-year overall survival rate of 85%-88% for stage I disease [ 9 ; 10 ]. However, patients with advanced OCCC have a worse prognosis [ 10 ; 11 ]. In a cohort of Memorial Sloan Kettering Cancer Center patients, the 3-year overall survival rate was 90% for stage I patients but fell to 53.3% for stage III patients and 29.6% for stage IV patients [ 7 ]. OCCC also shows greater chemoresistance compared with other subtypes [ 5 ; 12 ]. Thus, the current literature shows that complete gross resection, including lymph node dissection of nodes at the time of debulking surgery for comprehensive staging, results in the best outcomes in early stages [ 13 - 17 ]. Nevertheless, the decision regarding whether to perform upfront debulking surgery or to administer neoadjuvant chemotherapy followed by interval debulking surgery currently relies on the same criteria applied for the other primary ovarian/Mullerian malignancies, with trials and scores being conducted aiming to help in the process [ 18 ; 19 ]. From an operative point of view, since OCCC is more often associated with endometriosis, it is worth highlighting that it tends to display adherence to adjacent structures [ 20 ]. Furthermore, it is worth noting that at intraoperative diagnosis based on frozen sections, even though it is not considered especially crucial to distinguish OCCC from other entities as long as the Mullerian/primary ovarian origin is confirmed, frozen sections of OCCC tend not to have clear cytoplasm, and the section prepared during surgery may not demonstrate all characteristics necessary for diagnosis [ 21 ]; in turn, this may influence disease management in certain cases [ 22 ; 23 ]. Finally, significant molecular differences between OCCC and non-OCCC have been reported, with ARID1A/PIK3CA mutations being the most common in the former and P53 mutations being less common [ 5 ; 6 ]. At our institution, the gold standard for diagnosing ovarian clear cell carcinoma remains an expert gynecologic pathologist's careful evaluation of tumor morphology. For most of the remaining cases, a diagnosis of OCCC is generally favored on morphologic grounds and then confirmed by select immunohistochemistry. Regarding molecular studies, we do not typically use these as a diagnostic tool in OCCC as there is a longer turnaround time, and the results are generally not specific to this entity. Currently, surgery is the standard approach for the initial staging and treatment of OC. However, given the differences in treatment response and prognosis between OCCC and other subtypes, informing the surgeon of the potential for diagnosis of OCCC before surgery may add value. Indeed, awareness of the suspicion of OCCC could improve surgical planning by alerting the care team to potential challenges in resection and a possible need to involve other surgical specialists; furthermore, it may influence the decision regarding whether to administer neoadjuvant chemotherapy, given the relatively poor response of OCCC to platinum agents. Contrast-enhanced computed tomography (CECT) has repeatedly shown its value in the staging of OC, serving as a roadmap for surgical planning and being used in survival models [ 24 - 30 ]. We hypothesized that, in addition to having an established role in tumor staging for patients with OC, preoperative CT could help predict the diagnosis of OCCC. The purpose of our study was therefore to evaluate clinical and laboratory variables, as well as radiological variables from preoperative CECT, for their ability to distinguish OCCC from other OC subtypes (non-OCCC), and to develop nomogram models using these variables to preoperatively predict the probability of OCCC versus non-OCCC.