Vessel morphology depicted by three‐dimensional power Doppler ultrasound as second‐stage test in adnexal tumors that are difficult to classify: prospective diagnostic accuracy study

ABSTRACT Objectives To assess whether vessel morphology depicted by three‐dimensional (3D) power Doppler ultrasound improves discrimination between benignity and malignancy if used as a second‐stage test in adnexal masses that are difficult to classify. Methods This was a prospective observational international multicenter diagnostic accuracy study. Consecutive patients with an adnexal mass underwent standardized transvaginal two‐dimensional (2D) grayscale and color or power Doppler and 3D power Doppler ultrasound examination by an experienced examiner, and those with a ‘difficult’ tumor were included in the current analysis. A difficult tumor was defined as one in which the International Ovarian Tumor Analysis (IOTA) logistic regression model‐1 (LR‐1) yielded an ambiguous result (risk of malignancy, 8.3% to 25.5%), or as one in which the ultrasound examiner was uncertain regarding classification as benign or malignant when using subjective assessment. Even when the ultrasound examiner was uncertain, he/she was obliged to classify the tumor as most probably benign or most probably malignant. For each difficult tumor, one researcher created a 360° rotating 3D power Doppler image of the vessel tree in the whole tumor and another of the vessel tree in a 5‐cm3 spherical volume selected from the most vascularized part of the tumor. Two other researchers, blinded to the patient's history, 2D ultrasound findings and histological diagnosis, independently described the vessel tree using predetermined vessel features. Their agreed classification was used. The reference standard was the histological diagnosis of the mass. The sensitivity of each test for discriminating between benign and malignant difficult tumors was plotted against 1 – specificity on a receiver‐operating‐characteristics diagram, and the test with the point furthest from the reference line was considered to have the best diagnostic ability. Results Of 2403 women with an adnexal mass, 376 (16%) had a difficult mass. Ultrasound volumes were available for 138 of these cases. In 79/138 masses, the ultrasound examiner was uncertain about the diagnosis based on subjective assessment, in 87/138, IOTA LR‐1 yielded an ambiguous result and, in 28/138, both methods gave an uncertain result. Of the masses, 38/138 (28%) were malignant. Among tumors that were difficult to classify as benign or malignant by subjective assessment, the vessel feature ‘densely packed vessels’ had the best discriminative ability (sensitivity 67% (18/27), specificity 83% (43/52)) and was slightly superior to subjective assessment (sensitivity 74% (20/27), specificity 60% (31/52)). In tumors in which IOTA LR‐1 yielded an ambiguous result, subjective assessment (sensitivity 82% (14/17), specificity 79% (55/70)) was superior to the best vascular feature, i.e. changes in the diameter of vessels in the whole tumor volume (sensitivity 71% (12/17), specificity 69% (48/70)). Conclusion Vessel morphology depicted by 3D power Doppler ultrasound may slightly improve discrimination between benign and malignant adnexal tumors that are difficult to classify by subjective ultrasound assessment. For tumors in which the IOTA LR‐1 model yields an ambiguous result, subjective assessment is superior to vessel morphology as a second‐stage test. © 2020 The Authors. Ultrasound in Obstetrics & Gynecology published by John Wiley & Sons Ltd on behalf of International Society of Ultrasound in Obstetrics and Gynecology.


Vascular morphology in difficult adnexal masses
325 the ultrasound examiner was uncertain regarding classification as benign or malignant when using subjective assessment. Even when the ultrasound examiner was uncertain, he/she was obliged to classify the tumor as most probably benign or most probably malignant. For each difficult tumor, one researcher created a 360 • rotating 3D power Doppler image of the vessel tree in the whole tumor and another of the vessel tree in a 5-cm 3 spherical volume selected from the most vascularized part of the tumor. Two other researchers, blinded to the patient's history, 2D ultrasound findings and histological diagnosis, independently described the vessel tree using predetermined vessel features. Their agreed classification was used. The reference standard was the histological diagnosis of the mass. The sensitivity of each test for discriminating between benign and malignant difficult tumors was plotted against 1 -specificity on a receiver-operating-characteristics diagram, and the test with the point furthest from the reference line was considered to have the best diagnostic ability.

Conclusion Vessel morphology depicted by 3D power Doppler ultrasound may slightly improve discrimination between benign and malignant adnexal tumors that are difficult to classify by subjective ultrasound assessment. For tumors in which the IOTA LR-1 model yields an ambiguous result, subjective assessment is superior to vessel morphology as a second-stage test. © 2020 The Authors. Ultrasound in Obstetrics & Gynecology published by John Wiley & Sons Ltd on behalf of International Society of Ultrasound in Obstetrics and
Gynecology.

INTRODUCTION
Subjective assessment of ultrasound findings (also called pattern recognition) by an experienced ultrasound examiner is the best ultrasound method for discriminating between benign and malignant adnexal masses 1,2 .
However, even an experienced ultrasound examiner may find up to 10% of tumors impossible to classify confidently as benign or malignant using pattern recognition (termed 'difficult tumors') [3][4][5] . The 10% risk cut-off of the International Ovarian Tumor Analysis (IOTA) logistic regression model-1 (LR-1) has almost as good an ability to discriminate between benign and malignant tumors as has subjective assessment 6,7 , but it has been suggested that a risk of malignancy calculated by LR-1 of 8.3% to 25.5% represents an ambiguous risk 8 . The tumor marker CA 125 is clearly inferior to subjective assessment for discriminating between benign and malignant adnexal masses 9 and has no role in classifying difficult tumors [3][4][5]10 . Subjective assessment of ultrasound images is superior to computed tomography for discriminating between benign and malignant adnexal masses, while the role of magnetic resonance imaging is still unclear [11][12][13] . A logistic-regression model for calculating the risk of malignancy in tumors not classifiable as benign or malignant by an experienced ultrasound examiner using subjective assessment has been published, but its ability to discriminate between benign and malignant tumors was not superior to that of subjective assessment 4 .
For tumors that are difficult to classify as benign or malignant using subjective assessment or using the IOTA LR-1 model, a second-stage test capable of correctly classifying difficult tumors as benign or malignant would be valuable. A possible second-stage test is three-dimensional (3D) power Doppler ultrasound examination of the vascular tree of tumors 14 .
The aims of this study were to assess whether vessel morphology, as depicted by 3D power Doppler ultrasound, differs between benign and malignant difficult adnexal masses, and whether vessel morphology improves discrimination between benign and malignant masses when used as a second-stage test in difficult adnexal masses.

Study population
Our study population comprised those patients in the IOTA-3 study who had a difficult adnexal tumor (defined below) 15 . The IOTA-3 study is a prospective observational international multicenter cross-sectional diagnostic accuracy study (study protocol available in Appendix S1), which has been described in detail elsewhere 15 . Patients were recruited into IOTA-3 between October 2009 and May 2012 in 18 centers in six countries (Sweden, Belgium, Italy, Poland, Spain and the Czech Republic). These centers were either oncology referral centers (i.e. tertiary referral centers with a specific gynecological oncology unit) or other hospitals or units with a special interest in gynecological ultrasound. The centers and type of center are listed after the main text. Ethical approval was obtained from the Ethics Committee of the University Hospitals Leuven (B32220095331/S51375) as well as from the local ethics committees of all participating centers.

Sladkevicius et al.
Patients referred to one of the participating centers for an ultrasound examination and found to have an adnexal mass were eligible for inclusion in IOTA-3. Consecutive patients with at least one adnexal mass judged not to be a functional cyst, examined with transvaginal ultrasound by an experienced ultrasound examiner, were included in IOTA-3, provided that they gave written and/or oral informed consent before the ultrasound scan. If more than one mass was detected, the mass with the most complex ultrasound morphology was used for statistical analysis. When masses with similar morphology were observed, the largest mass or the one most easily accessible with ultrasound was used. Criteria for excluding patients from IOTA-3 were pregnancy at the time of the ultrasound examination, surgical removal of the mass more than 120 days after the ultrasound examination, data inconsistencies that persisted after final manual data checks and cases with incomplete histology.

Data collection
A dedicated, secure electronic data-collection system was developed for the IOTA-3 study (IOTA-3 Study Screen; astraia Software, Munich, Germany). Patients automatically received a unique identifier. Data security was ensured by encrypting all data communications. Data integrity and completeness were ensured by client-side checks in the system supplied by astraia and by final data cleaning by a group of biostatisticians and expert ultrasound examiners. The astraia software automatically calculated the risk of malignancy using the IOTA LR-1 model 6 , and the result was displayed on the computer screen.

Ultrasound examination
All patients included in the IOTA-3 study underwent a standardized transvaginal ultrasound examination by a gynecologist or radiologist who was highly experienced in gynecologic ultrasound 15 , using high-end ultrasound systems. Grayscale and color or power Doppler ultrasound was used to obtain information on more than 40 ultrasound variables to characterize each adnexal mass. The standardized ultrasound examination technique and the IOTA terminology used to describe the ultrasound images have been described elsewhere 16 . After completing the ultrasound examination, the examiner classified each mass as benign or malignant on the basis of his/her subjective assessment of the grayscale and color or power Doppler ultrasound findings. In addition, the examiner stated his/her level of confidence by classifying each mass as certainly benign, probably benign, uncertain, probably malignant or certainly malignant. This means that, even when the examiner was uncertain whether the tumor was benign or malignant, he/she was obliged to classify it as most probably benign or most probably malignant. The ultrasound information was entered prospectively into the electronic data-collection system (see above), and was locked at the time of the examination and could not be changed thereafter. The decision regarding surgery for adnexal tumors was made by the referring physician, based on clinical information, such as symptoms, age, operative risk and coexisting disease, and on the ultrasound report, which was written using the results of subjective assessment.
In addition to performing the two-dimensional (2D) ultrasound examination as described above, ultrasound examiners in centers with access to a Voluson 730 Expert or GE E8 ultrasound system (GE Healthcare, Zipf, Austria) with a 5-9-MHz or a 6-12-MHz vaginal transducer were asked to acquire 3D power Doppler ultrasound volumes of all adnexal masses. The power Doppler and 3D settings are described in Appendix S2. The examiners were instructed to include the whole tumor in the volume or, if this was not possible because the tumor was too large, to acquire several volumes to ensure that all parts of the tumor were captured. Patients were asked to lie still during the acquisition and, if necessary, to hold their breath. Volumes from difficult tumors, i.e. tumors in which the ultrasound examiner was uncertain whether the tumor was benign or malignant based on subjective assessment, or those in which the IOTA LR-1 model gave an ambiguous risk of malignancy (8.3% to 25.5%) 8 , were sent to Skåne University Hospital, Malmö, on compact discs, for analysis. The same methodology of analyzing the 3D volumes as described previously was used 14 , and is outlined briefly below.

Analysis of 3D volumes and audio-video interleave (AVI) files
For each tumor, 360 • rotating 3D power Doppler images (AVI-files) of the vessel tree in the whole tumor as well as in a 5-cm 3 spherical volume selected from the most vascularized part of the tumor were prepared by the second author (L.J.) using the Virtual Organ Computer-aided AnaLysis imaging program (4D-View, version 7.0, GE Healthcare) on a personal computer. Subjective assessment was used to select the most vascularized part of the tumor 14 . Before creating the rotating image of the vessel tree, color transparency was adjusted to optimize the delineation of the vessels.
All AVI files were then analyzed independently by two members of the Malmö research team (L.V., P.S.) who had no knowledge of patient history, 2D ultrasound findings or histological diagnosis. The vessel tree in the whole tumor volume, as well as in the 5-cm 3 sample, was characterized using the same classification as described previously ( Figure 1): branching, i.e. division of a vessel into two or more branches; changes in diameter, i.e. changes in vessel width from narrow to wide and from wide to narrow; 'splashes', i.e. areas of color in contrast to clearly separate vessels; tortuosity; areas with densely packed vessels; and 'bridges', i.e. straight connections between two nearby vessels. The presence of bridges was assessed only in the 5-cm 3 samples and the presence of densely packed vessels was assessed only in the whole tumor volume 14 .
In cases of disagreement between the two observers, consensus was reached by discussion, and the agreed classification was used for statistical analysis; this was done to decrease the risk of bias introduced by relying on one single observer. Ultrasound images of the vascular features are shown in Figures 2 and 3, and Videoclips S1-S7.

Reference standard
The reference standard was the histological classification of the excised mass as malignant or benign. Histological examination was carried out at the local centers. Central pathology review was not performed because, in a previous IOTA study, no clinically significant differences between local and central pathology reports were observed 6 . Malignant tumors were classified according to the criteria recommended by the International Federation  of Gynecology and Obstetrics 17 . Borderline ovarian tumors were classified as malignant. The pathologists were blinded to the ultrasound findings.

Statistical analysis
SAS version 9.4 (SAS Institute, Cary, NC, USA) was used for statistical analysis. The statistical significance of differences in proportions was determined using the χ-square test or Fisher's exact test and that of differences in continuous data using Student's t-test or the Mann-Whitney U-test, as appropriate.
We evaluated the discriminative ability of each vascular morphology feature, of subjective assessment and of the IOTA LR-1 model when using the recommended 10% risk cut-off for malignancy 6 , and expressed it as sensitivity, specificity and positive and negative likelihood ratios. The sensitivity of each test was plotted against 1 -specificity on a receiver-operating-characteristics (ROC) diagram; the test with the point furthest from the reference line was considered to have the best discriminative ability.
We estimated interobserver reliability (agreement beyond chance) in the assessment of the vascular tree of the tumors by calculating Cohen's κ 18 . κ values of 0.81-1.0 were taken to indicate excellent reliability, 0.61-0.80 good reliability and 0.41-0.60 moderate reliability 19 . P < 0.05 was regarded as statistically significant and was corrected for multiple testing using the permutation method 20 .

Sample-size calculation
We aimed to collect information on 300 difficult masses, of which we expected a minimum of 30% (n = 90) to be malignant 3,8 . Ninety malignancies would give us reasonable 95% CIs around the point estimates for sensitivity. Because about 7% of all adnexal masses are difficult to classify as benign or malignant using subjective assessment 4 , and because the IOTA LR-1 model may yield an ambiguous test result in 10% of all adnexal masses 8 , we estimated that we needed to examine about 2000 women with an adnexal mass (Appendix S1).
The study is reported using the Standards for Reporting Diagnostic Accuracy studies (STARD) guidelines 21 .

RESULTS
In total, 2541 women with an adnexal mass were enrolled for inclusion in IOTA-3, of whom 138 were excluded. Reasons for exclusion were > 120 days between ultrasound examination and surgery (n = 66), pregnancy (n = 31), data errors that could not be solved by contacting the principal investigators (n = 28) and incomplete final histology (n = 13). The final IOTA-3 dataset included 2403 patients 15 , of whom 376 (16%) had a difficult tumor; an uncertain result regarding malignancy was obtained for 168 (7%) tumors on subjective assessment, for 259 (11%) tumors using the LR-1 model and for 51 (2%) tumors using both methods. Serous and mucinous cystadenomas/cystadenofibromas, fibromas and borderline tumors were substantially more common among the difficult tumors than among the other ones, while endometriomas, benign teratomas (dermoid cysts), primary ovarian cancers and metastases in the ovaries from another primary tumor were substantially less common, with 2-to 3-fold differences in prevalence (Tables S1 and S2). The distribution of histological diagnoses was similar in tumors in which the ultrasound examiner was uncertain about the diagnosis on subjective assessment and in those in which LR-1 yielded an ambiguous result, with the exception that benign teratomas and endometriomas were more common in the latter (Table S1). Unilocular solid tumors, multilocular solid tumors and papillary projections were substantially more common in difficult tumors than in the others, while unilocular cysts, ground-glass echogenicity of cyst fluid, and color scores of 1 and 4 were substantially less common, with 2-to 3-fold differences in prevalence (Tables S3 and S4). The ultrasound characteristics of tumors that the examiner found difficult to classify on subjective assessment were similar to those in which LR-1 yielded an ambiguous result, with the exception that ascites was more common and unilocular cysts were less common in tumors that the examiner found difficult to classify on subjective assessment (Table S3).
3D ultrasound volumes were available for 138 of the 376 difficult tumors. Six centers did not provide volumes for any of their difficult masses (0/55), four centers provided volumes for more than 80% (73/87), four centers for between 40% and 57% (47/96), three centers for between 11% and 18% (16/107) and one center for 6% (2/31). In three centers, the reason for providing volumes for none or only a small proportion of the difficult tumors (6/78) was that the volumes stored in the ultrasound system had been deleted and there was no back-up, and in one center the GE ultrasound system required was not always available for research. One center (with 33 difficult tumors) did not have access to a Voluson 730 Expert or GE E8 ultrasound system and so could not provide any volumes. Seven centers reported forgetfulness or transient technical problems to be the explanation for not providing volumes for all their difficult masses, and six centers gave no explanation. The number of patients and the proportion of difficult tumors contributed by each center are shown in Table S5, and patient flow is described in Figure 4.
The available volumes from the 138 difficult tumors included 79 (57%) masses in which subjective assessment gave an uncertain result, 87 (63%) cases in which LR-1 gave an ambiguous result and 28 (20%) cases in which both methods gave an uncertain result. The histological diagnoses of the 138 difficult tumors with available volumes are shown in Table 1. One hundred (72%) tumors were benign and 38 (28%) were malignant. The distribution of histological diagnoses was similar to that in the 238 difficult tumors for which tumor volumes were not available, with the exception that the proportion of serous and mucinous cystadenomas/cystadenofibromas was slightly higher among the tumors with available volumes ( Figure 4, Table S6). The clinical background data and 2D ultrasound findings for the 138 difficult tumors with available volumes are described in Table 2. They were similar to those in the 238 difficult tumors for which volumes were not available, with the following exceptions: unilocular cysts, incomplete septa, tender mass on scan and color score of 1 were less common in the 138 difficult tumors for which tumor volumes were available, while multilocular solid tumors and tumors with color score of 3 or 4 were more common ( Figure 4, Table S7).
The ability of subjective assessment, the IOTA LR-1 model and the vascular features to discriminate between benign and malignant difficult tumors, in terms of sensitivity and specificity, is shown in Table 3 (95% CIs are shown in Table S8). All vessel features differed significantly between benign and malignant difficult tumors. Branching vessels, densely packed vessels, changes in diameter, tortuous vessels, color splashes and bridges between vessels were more common in the malignant than in the benign difficult tumors. However, none of the vessel features discriminated well between benign and malignant difficult tumors. Figure 5 shows plots of sensitivity against 1 − specificity for subjective assessment, for the IOTA LR-1 model when using a 10% risk cut-off and for the vessel features with the best discriminative ability. Subjective assessment was the best method for discriminating between benign and malignant masses in the total study population of 138 difficult masses, followed by densely packed vessels in the whole tumor volume and tortuous vessels in the tumor biopsy sample. Among the 79 tumors that were difficult to classify as benign or malignant using subjective assessment, densely packed vessels in the whole tumor volume and tortuous vessels in the tumor biopsy sample had the best discriminative ability. For those tumors in which the IOTA LR-1 model yielded an ambiguous result, subjective assessment had the best discriminative ability, while changes in diameter in the tumor biopsy sample had the second best.

Figure 4
Flowchart summarizing recruitment of patients with adnexal tumor that was difficult to classify as benign or malignant. Prevalence of histological diagnoses and ultrasound features that differed most between tumors that were and those that were not difficult to classify are shown. BOT, borderline ovarian tumor.

DISCUSSION
We have shown that vessel morphology, as seen on 3D power Doppler ultrasound, differs between benign and malignant difficult adnexal tumors. Branching vessels, changes in diameter, color splashes, tortuous vessels, densely packed vessels and bridges between vessels were more common in malignant than in benign difficult tumors. However, none of the vascular features discriminated well between benign and malignant difficult tumors. Our findings confirm that interobserver reliability with regard to vessel morphology depicted by 3D power Doppler is moderate to good 14 .
To the best of our knowledge, there are no published studies exploring the ability of the morphology of tumor vessels depicted by 3D power Doppler ultrasound to discriminate between benign and malignant difficult tumors. A strength of our study is that vessel morphology was assessed by observers who were blinded to clinical information, 2D grayscale or color Doppler ultrasound findings and histological diagnosis of the tumors. This means that our results regarding evaluation of the vessel tree are unbiased and reflect the true discriminative capacity of vessel morphology. A limitation of our study is possible selection bias, because not all the participating centers provided 3D volumes of all their difficult tumors. The histology and ultrasound features differed slightly between the difficult tumors included (i.e. those with tumor volumes available) and those not included (i.e. those with tumor volumes not available) in that the proportion of serous and mucinous cystadenomas/cystadenofibromas, multilocular solid tumors and tumors with a color score of 3 or 4 was higher among the difficult tumors that were included, while the proportion of unilocular cysts was lower (Figure 4, Tables S6 and S7). However, the histological diagnoses and ultrasound features of the difficult tumors with available volumes were fairly similar to those of all difficult tumors included in IOTA-3 (Table S1). Therefore, 14 (10) Data are given as mean ± SD, n (%), n/N (%) or median (range). *Data available for 117 cases.
our study sample should be reasonably representative of all difficult adnexal tumors. The small number of difficult masses with available volumes is another limitation, making our estimates of sensitivity and specificity imprecise (Table S8). However, because this was an exploratory study, we believe that this is acceptable. Our results confirm those of a previous study demonstrating that serous and mucinous cystadenomas/cystadenofibromas, fibromas and borderline tumors are difficult to classify as benign or malignant, and that unilocular solid and multilocular solid tumors and tumors with papillary projections are over-represented among difficult tumors 4 . The proportion of tumors that the ultrasound examiner found difficult to classify based on subjective assessment in the current study is the same as that in the previously published study (168/2403 (7%) vs 244/3511 (7%)) 4 .
To the best of our knowledge, the ability of vessel morphology depicted by 3D ultrasound to discriminate between benign and malignant adnexal masses or to decrease diagnostic errors has been explored in only one published study 14 , which included 104 adnexal masses reasonably representative of a general population of tumors scheduled for surgery. Similar to our findings, all   Our results show that, if the IOTA LR-1 model gives an ambiguous risk estimate (8.3% to 25.5%), then subjective assessment by an experienced ultrasound examiner is superior to using vessel morphology depicted by 3D power Doppler ultrasound as a second-stage test. However, if a mass cannot be confidently classified as benign or malignant by an experienced ultrasound examiner using subjective assessment, then assessing vessel morphology with 3D power Doppler ultrasound could be useful. Both densely packed vessels in the whole tumor volume and tortuous vessels in a 5-cm 3 tumor biopsy can be used for discrimination ( Figure 5). We recommend using densely packed vessels because interobserver reliability was better for this variable than for tortuous vessels in a tumor biopsy. On the other hand, it is more time consuming to analyze a whole tumor volume than a tumor biopsy 22 . In our experience, for a highly experienced ultrasound examiner, it takes a minimum of 2.5 min to create a rotating 3D image of the vascular tree of a whole tumor, while it takes a minimum of 1 min to create one for a 5-cm 3 biopsy selected from the most vascularized part of the tumor. When assessing vessel morphology, it is important to be aware of the pitfalls of Doppler ultrasound and to ensure that Doppler settings are correct. If the tumor is far away from the ultrasound probe, it might not be possible to detect Doppler signals from the whole or parts of the tumor. This limits the clinical usefulness of vessel morphology for classifying tumors as benign or malignant. Another limiting factor is the subjectivity of the method. Evaluation of vessel morphology, including 'densely packed vessels', is based on pattern recognition and is therefore difficult to standardize or define.
Correct classification of adnexal masses as benign or malignant is a requirement for optimal management, i.e. conservative management with follow-up examinations, surgery in a local hospital or referral to a center specialized in gynecological oncology 23 . However, some tumors are difficult to classify confidently as benign or malignant. Vessel morphology as depicted by 3D power Doppler ultrasound shows limited ability to discriminate between benign and malignant difficult tumors. It remains to be seen whether new biomarkers, immune cells, proteins or genetic information can improve the classification of difficult tumors as benign or malignant.

ACKNOWLEDGMENTS
This study was supported by the Swedish Medical Research Council (grants no. K2006-73X-11605-11-3); by the Swedish state under the agreement between the Swedish government and the county councils, the ALF-agreement; Landstingsfinansierad regional forskning (a Swedish governmental grant from the region of Scania); funds administered by Skåne University Hospital; and Allmänna Sjukhusets i Malmö Stiftelse för bekämpande av cancer (the Malmö General Hospital Foundation for fighting against cancer).

SUPPORTING INFORMATION ON THE INTERNET
The following supporting information may be found in the online version of this article: Appendix S1 International Ovarian Tumor Analysis (IOTA) Phase 3 study protocol Appendix S2 Power Doppler and three-dimensional (3D) ultrasound settings used in study Videoclips S1-S7 Three-dimensional 360 • rotating power Doppler ultrasound images of vessel tree in: benign mucinous cystadenoma, showing dispersed (as opposed to densely packed), straight (as opposed to tortuous), branching vessels (Videoclip S1; still image of same tumor is Figure 2a Tables S1 and S2 Histological diagnoses of 2403 adnexal tumors, according to: whether tumor was difficult to classify and availability of ultrasound volumes (Table S1) and whether tumor was difficult to classify as benign or malignant (Table S2) Tables S3 and S4 Clinical and ultrasound characteristics of 2403 adnexal tumors, according to: whether tumor was difficult to classify and availability of ultrasound volumes (Table S3) and whether tumor was difficult to classify as benign or malignant (Table S4)