To determine whether subjective evaluation of the morphology of the vessel tree of ovarian tumors, as depicted by three-dimensional (3D) power Doppler ultrasound, can discriminate between benign and malignant ovarian tumors, and whether it improves characterization compared with using gray-scale ultrasound imaging alone.
A consecutive series of 104 women scheduled for surgical removal of an ovarian mass were examined with transvaginal two-dimensional (2D) gray-scale and 3D power Doppler ultrasound. Predetermined vessel characteristics, e.g. density of vessels, branching, caliber changes and tortuosity, were evaluated in 360° rotating 3D images of the vessel tree of the tumor. Ultrasound results were compared with those of the histology of the surgical specimens. Univariate and multivariate logistic regression were used.
There were 77 benign tumors, six borderline tumors and 21 invasive malignancies. All vascular features differed significantly between benign and malignant tumors. The areas under their receiver–operating characteristics (ROC) curves (AUCs) were in the range 0.61–0.83. The AUC of a logistic regression model containing three gray-scale ultrasound variables was 0.98. This model correctly classified all malignancies, with a false-positive rate of 10% (8/77). Adding branching of vessels in the whole tumor to the gray-scale model yielded an AUC of 0.99 and resulted in all malignancies and an additional four benign tumors being correctly classified.
Both benign and malignant tumors are composed of ‘tumor cells’ and supporting tissues: blood vessels, fibrous supporting tissue, lymphatic vessels and sometimes nerve fibers1, 2. An increase in the cell population of tumors must be preceded by the production of new vessels—angiogenesis3. Usually newly formed vessels pierce the tumor from various directions from the surrounding tissue of the organ. Tumor vascular networks are tortuous and do not follow the regular structural hierarchy seen in normal tissue4. Vessels in malignant tumors are not well differentiated, manifest heterogeneity of structure5, are often densely distributed, dilated and saccular and may contain tumor cells within their endothelial lining6, 7. Malignant tumors may also harbor giant capillaries and arteriovenous shunts without intervening capillaries7, 8. Such haphazard branching patterns and larger, less regular diameters of vessels in tumors contribute to the non-uniform perfusion of cancer cells, expressed as chaotic tumor blood flow, with high flow rates in some vessel segments and stagnation in others7. These patterns can change within hours or even minutes9.
Tumor vascularity has been examined by microscopy10, 11, polymer casting techniques12, 13, contrast angiography14–16, color and power Doppler ultrasound17–24 and using ultrasound contrast25. Both two-dimensional (2D) and three-dimensional (3D) power Doppler ultrasound has been used to try to quantify the color content of tumor scans objectively19, 21, 26. Using 3D power Doppler ultrasound, a 3D image of the vessel tree of tumors can be created. However, we know of no publications providing a systematic description of the morphology of vessels in ovarian tumors, as assessed by 3D power Doppler ultrasound. Testa and coworkers27 briefly mention, in a study on computerized quantitative analysis of the color content of tumor volumes obtained by 3D ultrasound, that the ‘presence of irregular and randomly dispersed vessels with complex branching was considered suggestive of malignancy’. They found this vessel pattern to be more common in malignant than in benign tumors, but the difference was not statistically significant. The reproducibility of assessment of the vessel tree was not tested27.
The aims of this study were: to determine whether subjective evaluation of the morphology of the vessel tree of ovarian tumors, as depicted by 3D power Doppler ultrasound, is reproducible; whether the vascular morphology, as assessed by 3D power Doppler ultrasound, differs between benign and malignant ovarian tumors; and whether 3D power Doppler ultrasound adds anything to gray-scale ultrasound imaging in an ordinary population of tumors.
Subjects and methods
The study protocol was approved by the Ethics Committee of the Medical Faculty of Lund University, Sweden. Informed consent was obtained from all participants, after the nature of the procedures had been fully explained.
A consecutive series of 131 women scheduled for surgical removal of an adnexal mass that clinically and at ultrasound examination was judged to be of ovarian origin were examined using 3D transvaginal ultrasound. Exclusion criteria were: unequivocal ultrasound diagnosis on the basis of pattern recognition28 of tubal disorder (e.g. hydrosalpinx); peritoneal pseudocyst, paraovarian cyst or dermoid cyst; operation > 90 days after the ultrasound examination; or Doppler artifacts or other technical problems that made evaluation of the vascular tree of tumors unreliable. The same series of patients was used in a study evaluating the quantitative analysis of power Doppler signals in tumor volumes obtained by 3D ultrasound26.
Before the ultrasound examination a clinical history was taken from each patient, following a standardized research protocol. A woman was considered to be postmenopausal if she reported a period of at least 12 months of amenorrhea after the age of 40 years, provided that medication or disease did not explain the amenorrhea. Women 50 years or older who had undergone hysterectomy were also defined as postmenopausal.
The last author, who had 16 years of experience of gynecological ultrasound, carried out all ultrasound examinations. The women were examined in the lithotomy position with an empty bladder. The equipment used was a GE Voluson 730 Expert ultrasound system (GE Healthcare, Zipf, Austria) with a 2.8–10-MHz transvaginal transducer. Identical fixed pre-installed power Doppler ultrasound settings were used: frequency, 6–9 (‘normal’) MHz; pulse repetition frequency, 0.6 kHz; gain, − 4.0; wall motion filter, ‘low 1’ (40 Hz).
A standardized examination technique and standardized definitions of gray-scale ultrasound terms were used as described previously29. Measurements were taken using calipers on the frozen ultrasound image. A papillary projection was defined as any solid protrusion into a cyst cavity from the cyst wall with a height ≥ 3 mm29. The size of the lesion and that of its largest solid component was calculated as the mean of three orthogonal diameters. In cases of bilateral masses, data from the mass with the most complex gray-scale ultrasound morphology were used in our statistical analyses.
After completion of the gray-scale ultrasound examination, the ultrasound system was switched into the power Doppler mode and then into the 3D mode. Attempts were made to include the whole tumor, or as much of it as possible, in the 3D ultrasound volume. If the whole tumor could not be included in the volume, care was taken to include the most vascularized parts of the tumor. The woman was asked to remain still during acquisition of the volume. After acquisition, the resultant multiplanar view was examined to ensure that a complete volume of the ovarian tumor, or as large a part as possible of it, had been captured. Volumes of satisfactory quality were stored on compact disks for future analysis.
The second author carried out the analyses of stored ultrasound volumes. The analyses were done off-line, using the Virtual Organ Computer-aided AnaLysis (VOCAL™) imaging program (4D-VIEW, version 2.1; GE Healthcare) on a personal computer. The acquired volumes yielded multiplanar views of the ovarian tumor in the midsagittal, transverse and coronal planes. The longitudinal plane through the tumor was used as the reference image. The rotation steps were 30°, resulting in the definition of six contours of the ovarian tumor. Contours of the tumor were manually drawn in all six sections, using the computer mouse. A volume was calculated automatically using the six contours drawn, and the color rendering mode was selected to visualize the vessel tree within the volume so defined. A 360° rotating 3D image of the vessels tree was obtained using the cine rotation function and saved as an AVI file for later analysis, as described below. A 5-cm3 spherical volume, called the ‘sample’, was selected from that part of the tumor volume that appeared to be most vascularized at visual inspection. A 3D image of the vessel tree in the sample was created and saved as described above for the whole tumor. The size of the sample (5 cm3) was chosen arbitrarily; it would have been difficult to assess the vascular tree in samples smaller than 5 cm3, and the smallest tumors in our study would not have filled a larger sample volume.
Approximately 12 months after the ultrasound examinations had been carried out, the first and third authors, who both had more than 15 years of experience in gynecological ultrasound, independently analyzed the 360° rotating 3D images of the vessel tree of the whole tumor and of the selected 5-cm3 sample. On a dedicated paper form they noted the presence or absence of a number of predefined vascular features. These had been chosen partly on the basis of theoretical knowledge of vessel morphology in malignant tumors4–8 and partly on the basis of previous personal experience of looking at vessel trees as depicted by 3D power Doppler ultrasound of normal organs and tumors. The following vessel features were studied: branching, i.e. division of the vessel into two or more branches; caliber changes, i.e. changes in vessel width from narrow to wide and again from wide to narrow; ‘splashes’, i.e. irregular areas of color in contrast to clearly separate vessels; tortuosity; and ‘bridges’, i.e. straight vessel connections between two nearby vessels. The presence of bridges was assessed only in the 5-cm3 samples, because of difficulties with detecting bridges in large volumes. In addition, to give an overview of tumor vascularization as a whole, the distribution of vessels within the whole tumor was described as: ‘dense’, i.e. the whole tumor filled with densely packed vessels; ‘dense areas’, i.e. having one or more areas where vessels were densely packed; or ‘not dense’, i.e. with widely dispersed vessels. Schematic drawings and power Doppler images of the vascular features are shown in Figures 1–3 and in the on-line Supplementary files (AVI files, Figures S1 and S2). After having independently assessed all AVI files, the observers reviewed their filled-in paper forms together. If there was disagreement for any of the predefined vessel variables, they went back to the AVI files and reviewed them together to reach a consensus. Their agreed classification was used for further statistical analysis.
The results of the ultrasound examinations were compared with those of the histological examination of the respective surgical specimens. Staging of malignant tumors was done by the attending physician, in accordance with the classification system recommended by the International Federation of Gynecology and Obstetrics30. In the statistical calculations, borderline tumors were classified as malignant.
Statistical calculations were undertaken using the Statistical Package for the Social Sciences (SPSS Inc., Chicago, IL, USA, version 12.02).
Interobserver agreement in the assessment of the vascular tree was determined by calculating Cohen's kappa (κ) value31, κ values of 0.81–1.0 indicating excellent agreement, 0.61–0.80 good agreement and 0.41–0.60 moderate agreement32.
The statistical significance of a possible relationship between malignancy and the features of the vascular tree of the tumor was determined using univariate logistic regression with the likelihood ratio test. Multivariate logistic regression was used to determine the effect of adding information on the vessel tree to a logistic regression model, including three gray-scale ultrasound variables: the size of the largest solid component of the tumor; irregular wall; and lesion size. The gray-scale model designed to calculate the risk of malignancy in an adnexal mass was built using the same patients as in the current study, and has been described in a previous publication26. To avoid overfitting, only one vascular feature was allowed to enter the gray-scale model, i.e. a maximum of four variables were allowed in the new model. The likelihood ratio test was used to determine which variables to include in the logistic regression model, P < 0.05 being the threshold for inclusion.
The application of the regression equations to data from each woman gave the probability of that woman having a malignant tumor (range 0–1). Receiver-operating characteristics (ROC) curves33 were drawn for single predicting variables and for regression equations. The area under the ROC curve and the 95% confidence interval (CI) of this area were calculated. If the lower limit of the CI for the area under the ROC curve was > 0.5, the diagnostic test was considered to have a discriminatory potential. The ROC curves were also used to determine the mathematically best cut-off value for each diagnostic test (single variables as well as logistic regression models). The mathematically best cut-off value corresponded to the point on the ROC curve situated farthest away from the reference line33. The sensitivity, false-positive rate and positive and negative likelihood ratios (LR) with regard to malignancy of the mathematically best cut-off values were also calculated. The model yielding the largest area under the ROC curve, the highest positive LR and the lowest negative LR for the mathematically best cut-off value was considered to be the best test. Two-tailed values of P⩽0.05 were considered statistically significant.
Of the 131 consecutive women examined, 24 were excluded because of an unequivocal ultrasound diagnosis of dermoid cyst (n = 14), tubal disease (n = 7), paraovarian cyst (n = 2) or peritoneal pseudocyst (n = 1). The ultrasound diagnosis was correct in 23 of these 24 cases (one ultrasound diagnosis of dermoid cyst was incorrect, the histopathology of the tumor being endometrioma). Three women were excluded because of technical problems or power Doppler artifacts that made evaluation of the vascular tree of the tumor unreliable. Thus, 104 tumors were included.
Among these 104 tumors there were 77 (74%) benign tumors and 27 (26%) malignancies. Histological diagnoses are presented in Table 1. The mean ( ± SD) age of the women with a benign tumor was 45 ± 17.4 years, and the mean age of those with a malignant tumor was 56 ± 17.6 years; 35% (27/77) and 56% (15/27), respectively, were postmenopausal; and 30% in both groups (23/77 vs. 8/27) were nulliparous. The most important gray-scale ultrasound features of the tumors are shown in Table 2. Power Doppler signals were detected in all tumors.
Table 1. Histopathological diagnoses
Fourteen epithelial cancers (six serous adenocarcinomas, three endometrioid adenocarcinomas, four adenocarcinomas of low differentiation and one clear cell cancer mixed with cancer of low differentiation), one granulosa cell tumor, one sex cord stromal cell tumor, one malignant teratoma and one malignant mixed Mullerian tumor.
Interobserver agreement with regard to the assessment of the vascular tree was moderate to good (κ = 0.44–0.78; Table 3). The highest κ values were achieved for density of vessels (κ = 0.70) and branching of vessels in the whole tumor (κ = 0.70) and for bridges between vessels (κ = 0.78) and color splashes (κ = 0.76) in the 5-cm3 sample.
Table 3. Interobserver agreement in the assessment of the morphology of the vascular tree of tumors
Sample from the most vascularized
area of the tumor
The diagnostic performance of the features of the vascular tree is summarized in Table 4. All vascular features differed significantly between benign and malignant tumors. Of the malignant tumors, 48% (13/27) had dense vessels in the whole tumor vs. 5% (4/77) of the benign tumors, 30% (8/27) vs. 10% (8/77) had areas with dense vessels, and 22% (6/27) vs. 85% (65/77) had widely dispersed vessels (P < 0.0001). The diagnostic performance of the features of the vascular tree was at most moderate. The areas under the ROC curves varied between 0.61 and 0.83, the largest areas being 0.83 for density of vessels in the whole tumor and 0.77 for bridges between vessels in the 5-cm3 sample. Branching of vessels in the 5-cm3 sample did not seem to have any discriminative power at all, the lower 95% confidence limit of the area under the ROC curve being 0.49. The strongest predictor of malignancy was the presence of dense vessels in the whole tumor, which increased the odds of malignancy five-fold. The strongest predictor of benignity was the absence of branching vessels in the whole tumor, which decreased the odds of malignancy 10-fold. The vascular features in the 5-cm3 sample did not change the likelihood of malignancy at all or only little, the positive LRs varying between 1.3 and 3.8 and the negative LRs between 0.3 and 0.7.
Table 4. Diagnostic performance of the morphology of the vascular tree
All vascular features except splashes in the whole tumor and branching in the 5-cm3 sample added information to the logistic regression model containing only gray-scale variables. The performance of the models is shown in Table 5. The gray-scale model correctly identified all 27 malignancies and 69 of the 77 benign tumors. One of the best models was obtained by adding branching of vessels in the whole tumor to the gray-scale model. The mathematical formula of this model is shown in the footnote to Table 5. Its area under the ROC curve was 0.99, with all malignancies and 73 of the 77 benign tumors being correctly classified, i.e. an additional four benign tumors were correctly classified compared to when using the gray-scale model without any added Doppler variable (comparison based on the use of the optimal cut-offs of both models).
Table 5. Diagnostic performance of logistic regression models
Higher values than the cut-off value indicate malignancy.
False-positive rate defined as 1 minus specificity.
Probability of malignancy = [ez/(1 + ez)], where z = (0.110 × mean diameter of largest solid component (mm)) + (0.028 × mean lesion diameter (mm)) + (2.15 × wall irregularity coded as 0 or 1) − 7.671, where e is the mathematical constant and base value of natural logarithms.
Probability of malignancy = [ez/(1 + ez)], where z = (0.147 × mean diameter of largest solid component (mm)) + (0.042 × mean lesion diameter (mm)) + (2.547 × wall irregularity coded as 0 or 1) + (6.100 × branching of vessels in the whole tumor coded as 0 or 1) − 15.447, where e is the mathematical constant and base value of natural logarithms. LR+, positive likelihood ratio; LR−, negative likelihood ratio; ROC, receiver–operating characteristics.
Depiction of the morphology of tumor vessels using 3D power Doppler ultrasound may be regarded as a new ultrasound modality. We wanted to determine whether subjective evaluation of the ultrasound morphology of tumor vessels, as depicted by 3D power Doppler ultrasound, is reproducible, and whether the morphology of the vessel tree so evaluated contains any clinically useful information. It is not possible to describe the morphology of the vessel tree using 2D color or power Doppler ultrasound. The only vascular feature that can possibly be evaluated using 2D ultrasound is vessel density, subjective evaluation of the color content of a 2D tumor scan probably being related to vessel density in a 3D power Doppler image of the vessel tree. However, the two are not directly comparable. The area under the ROC curve of vessel density, as assessed by 3D ultrasound in the current series of tumors, was similar to that of the color content of the tumor scan as evaluated using 2D ultrasound in (virtually) the same series of tumors26 (area under the ROC curve 0.83 vs. 0.80). It is interesting to note that the density of vessels was the single best vascular predictor of malignancy, because the density of vessels is likely to be reflected in the total color content of the tumor scan, and the color content of the tumor scan has been shown to be one of the best 2D Doppler ultrasound variables for discriminating between benign and malignant adnexal masses17. Others, too, have found, using methods other than ultrasound, that vessels are more densely packed in malignant than in benign tumors34–36.
Our results show that subjective evaluation of the morphology of the vessel tree of ovarian tumors, as depicted by 3D power Doppler ultrasound, manifest moderate to good interobserver reproducibility. The vessel features most difficult to reproduce were caliber changes and tortuosity. We have also shown that the morphology of the vascular tree of benign and malignant ovarian tumors, as depicted by 3D power Doppler ultrasound, does differ. This was perhaps to be expected, because the vessels of benign and malignant tumors have been described to be different when using methods other than ultrasound to describe them15, 16, 37. On the other hand, our study could just as well have shown that the 3D power Doppler method was too crude to detect the differences. Earlier studies found a wide heterogeneity of vessel distribution in different parts of tumors38, 39 and in different types of tumor37, 40, 41. Such heterogeneity may explain some of the false-positive and false-negative diagnoses using our technique.
In all previous studies evaluating the value of adding Doppler ultrasound information to gray-scale imaging, gray-scale ultrasound findings were available to the person evaluating the Doppler information because it was technically impossible to conceal it28, 42–48. Therefore, in the studies cited, Doppler results may have been biased. Because it was the aim of our study to determine the diagnostic performance of vessel morphology as depicted by 3D ultrasound, it is a methodological strength that vessel morphology was evaluated by observers having no clinical information, no information on histology, and no information on gray-scale ultrasound morphology. The AVI files contained no gray-scale information and had been prepared by a third person. Thus, our results of the evaluation of the vessel tree are completely unbiased and reflect the true capacity of vessel morphology to discriminate between benign and malignant adnexal masses.
Potential weaknesses of our study are that the very largest tumors could not be included in their entirety in the volume acquired, and that the site of the 5-cm3 sample was chosen on the basis of subjective evaluation of which area of the tumor was most vascularized. We do not believe that a small part (5–10% in most cases) of the very largest tumors missing invalidates our results, or that choosing the place of the 5-cm3 sample using subjective evaluation does. Any selection of a ‘representative’ part of a tumor must necessarily be subjective. Alcazar et al.19 used subjective evaluation to find the most ‘suspicious’ part of a tumor in a volume acquired by 3D ultrasound and quantified the color content of that part using the VOCAL™ software.
It is a weakness of our study that the logistic regression models including vascular morphology were not tested prospectively on a test set. However, it would be of limited interest to test our risk calculation models, including vessel morphology variables, prospectively in an ordinary tumor population because the gray-scale model itself performed extremely well in our study population (area under ROC curve 0.98), and it also performed very well (area under ROC curve 0.89) when tested prospectively in a test population of more than 1000 patients49, both study populations representing ordinary tumor populations. Five42–46 of seven studies42–48 evaluating prospectively whether adding Doppler ultrasound examination to gray-scale imaging improved discrimination between benign and malignant masses showed that the contribution of Doppler ultrasound examination to a correct diagnosis was very limited. In concordance, the results of our study on the vessel tree of tumors suggest that, in an ordinary population of ovarian tumors, 3D Doppler ultrasound examination adds little to gray-scale imaging. An experienced ultrasound examiner can usually correctly and confidently discriminate between benign and malignant adnexal tumors on the basis of gray-scale imaging alone, using pattern recognition28, 45. However, approximately 10% of adnexal tumors are difficult to classify correctly as benign or malignant, even for an experienced examiner45, 50. The difficult tumors are often borderline tumors, papillary cystadeno(fibro)mas or struma ovarii, and at ultrasound examination they are often seen to have papillary projections or to be multilocular cysts with a very large number of locules50. It is for these types of difficult tumor that gray-scale imaging needs to be supplemented by additional diagnostic methods. It would be interesting to test the value of adding evaluation of the morphology of the vascular tree, as assessed by 3D power Doppler ultrasound, to gray-scale imaging in a very large series of difficult tumors. The problem is that such tumors are rare, and it would require a large multicenter study to collect enough data within a reasonable time. One could design such a study as a randomized trial, where women with difficult tumors were assigned to have either a 2D gray-scale ultrasound examination alone, or a 2D gray-scale examination supplemented by conventional 2D color or power Doppler ultrasound, or by 3D power Doppler examination of the morphology of the vessel tree, and then to compare the diagnostic performance of the three strategies.
This study was supported by the Swedish Medical Research Council (Grant Nos K2001-72X-11605-06A, K2002-72X-11605-07B, K2004-73X-11605-09A and K2006-73X-11605-11-3), two governmental grants, Landstingsfinansierad Regional Forskning (Region Skåne and ALF-Medel) and funds administered by Malmö University Hospital.