7
The use of mathematical models to evaluate pelvic masses; can they beat an expert operator?

https://doi.org/10.1016/j.bpobgyn.2003.09.009Get rights and content

Abstract

The pre-operative characterization of ovarian cysts remains a major challenge. Functional cysts and some other benign cysts should be managed conservatively, whereas persistent tumours may need removal. It is crucial to distinguish between malignant tumours, which are better treated by a gynaecological oncologist, and benign tumours, which may be suitable for minimal-access surgery. Over the past decade several ultrasound-based morphological scoring systems, colour Doppler parameters, logistic regression models and artificial neural networks have been proposed and tested in order to try to predict the histology of ovarian tumours. On prospective testing none of the current models can beat an expert sonologist. Signs of malignancy include the presence of papillary structures, irregular solid areas, septa and a strong vascularization at colour Doppler imaging. Further refinement of mathematical models and the results of multicentre trials need to be reviewed before the clinical use of mathematical models can be advocated.

Section snippets

Serum CA 125 levels

The CA 125 antigen is a glycoprotein with a high molecular weight that is expressed by most epithelial ovarian cancers and is recognized by a monoclonal antibody (OC 125). Serum CA 125 is the tumour marker with the highest sensitivity for ovarian cancer.1., 2., 3., 4. This tumour marker will detect nearly 80% of advanced (stage ≥III) ovarian cancers, but only 40–44% of patients with stage I disease.5., 6., 7., 8. Some authors use different cut-off levels for pre- and post-menopausal patients.

Risk of malignancy index (RMI)

In order to increase the reliability of serum CA 125 levels to differentiate pre-operatively between benign and malignant pelvic masses, Jacobs and colleagues combined the measurement of CA 125 values with morphological findings and menopausal status of the patient in calculating a risk of malignancy index.15 In a retrospective study using transabdominal ultrasonography the following features suggestive of malignancy have been assessed: multiloculated cysts, evidence of solid areas, evidence of

Morphological scoring systems

More than 10 different morphological scoring systems have been reported and used with varying success.24., 25., 26., 27., 28., 29., 30., 31., 32., 33. Morphological features of pelvic masses are more extensively discussed in Chapter 6 and in other textbooks.34 A summary of clinically useful features is given in Table 3.35

Multivariate logistic regression analysis

Multivariate logistic regression analysis is a statistical tool that can be used to select and combine input variables which are linked to a certain outcome, for example, patient or tumour characteristics that are linked to the presence of malignancy in a pelvic mass. In the logistic regression model, the numerical values x1 to xn associated with observations of selected variables are weighted by coefficients β1 to βn and then summed together. An intercept β0 is subtracted after which the

Artificial neural network (ANN)

The use of artificial neural networks can be seen as a generalization of the methodology of logistic regression analysis described above. Here also an a posteriori probability of malignancy is modelled, but in an ANN the separation in the feature-space is more complex and non-linear in nature.

Artificial neural networks are networks of units (called neurons) that exchange information with each other in the form of numerical values via synaptic interconnections. The neurons take a weighted sum of

New statistical techniques

New statistical models are being developed and tested prospectively. Least squares support vector machine (LS-SVM) classifiers are known to have good generalization performance.45 Graphical models or Bayesian belief networks for grey-box model fitting offer the possibility to integrate expert knowledge, to estimate wrong or missing data, to derive confidence intervals, and to predict subclasses (e.g. borderline tumours, dermoid cysts).46., 47. However, so far these new techniques have not been

New ultrasound-based techniques

Several new ultrasound-based techniques show promise in distinguishing between malignant and benign adnexal tumours: for example, the kinetics of ultrasound contrast agents and three-dimensional power Doppler ultrasound. Orden et al measured uptake and washout times and demonstrated that after microbubble contrast injection malignant and benign adnexal lesions behave differently in degree, onset, and duration of Doppler ultrasound enhancement.48 Further prospective studies are needed to explore

New tumour markers and proteomic patterns

New tumour markers and developments in proteomic patterns50 are beyond the scope of this chapter, but it might be anticipated that quantitative results of these technologies could become incorporated in newly developed statistical models.

Subjective assessment of ultrasound images

One of the first prospective studies assessing the accuracy of subjective impressions of adnexal masses was performed between 1981 and 1985.51 Using transabdominal ultrasonography, an overall accuracy of 91% was obtained (see Table 5). The authors regarded thick septa, irregular solid parts within a mass, indefinite margins, and the presence of ascites and matted bowel loops as malignant patterns.

Using transvaginal grey-scale imaging subjective assessment, Valentin obtained an accuracy of 95%

Prospective comparison of methods

On internal prospective validation all mathematical models performed less well.37 The artificial neural network was significantly better than the three logistic regression models described above, but there was no significant difference between the performance of RMI and artificial neural networks (Timmerman et al, unpublished data).

As expected, external validation usually gives even more variable results in the estimates of diagnostic efficacy for malignancy of mathematical models. Aslam et al

Summary

An ANN can be trained to predict malignancy from the patient's age, CA125 levels, and some simple ultrasonographic criteria with a high degree of accuracy. The most important step in designing a logistic regression or a neural network model is cross-validation. In cross-validation, the ability of a trained network to generalize is evaluated by observing how the network performs on facts in the database that were withheld from the training set.63 Current mathematical models seem not to offer

References (71)

  • S Granberg et al.

    Macroscopic characterisation of ovarian tumors and the relation to the histological diagnosis: criteria to be used for ultrasound evaluation

    Gynecologic Oncology

    (1989)
  • S Granberg et al.

    Tumors in the lower pelvis as imaged by transvaginal sonography

    Gynecologic Oncology

    (1990)
  • J.P Lerner et al.

    Transvaginal ultrasonographic characterisation of ovarian masses with an improved, weighted scoring system

    American Journal of Obstetrics and Gynecology

    (1994)
  • P DePriest et al.

    Transvaginal sonography as a screening method for the detection of early ovarian cancer

    Gynecologic Oncology

    (1997)
  • L Valentin

    Gray scale sonography, subjective evaluation of the color Doppler image and measurement of blood flow velocity for distinguishing benign and malignant tumors of suspected adnexal origin

    European Journal of Obstetrics and Gynecology

    (1997)
  • C Lu et al.

    Preoperative prediction of malignancy of ovarian tumors using least squares support vector machines

    Artificial Intelligence in Medicine

    (2003)
  • P Antal et al.

    Bayesian applications of belief networks and multilayer perceptrons for ovarian tumor classification with rejection

    Artificial Intelligence in Medicine

    (2003)
  • L.S Cohen et al.

    Three-dimensional power Doppler ultrasound improves the diagnostic accuracy for ovarian cancer prediction

    Gynecologic Oncology

    (2001)
  • D Boll et al.

    The pre-operative assessment of the adnexal mass: the accuracy of clinical estimates versus clinical prediction rules

    British Journal of Obstetrics and Gynaecology

    (2003)
  • N Aslam et al.

    Prospective evaluation of logistic regression models for the diagnosis of ovarian cancer

    Obstetrics and Gynecology

    (2000)
  • E Ekerhovd et al.

    Preoperative assessment of unilocular adnexal cysts by transvaginal sonography: a comparison between sonographic morphological imaging and histopathologic diagnosis

    American Journal of Obstetrics and Gynecology

    (2001)
  • B Gerber et al.

    Simple ovarian cysts in premenopausal patients

    International Journal of Gynaecology and Obstetrics

    (1997)
  • C Bailey et al.

    The malignant potential of small cystic ovarian tumors in women over 50 years of age

    Gynecologic Oncology

    (1998)
  • E Kroon et al.

    Diagnosis and follow-up of simple ovarian cysts detected by ultrasound in postmenopausal women

    Obstetrics and Gynecology

    (1995)
  • I Vergote et al.

    Prognostic importance of degree of differentiation and cyst rupture in stage I invasive epithelial ovarian carcinoma

    Lancet

    (2001)
  • R.C Bast et al.

    A radioimmunoassay using a monoclonal antibody to monitor the course of epithelial ovarian cancer

    New England Journal of Medicine

    (1983)
  • I Jacobs et al.

    The CA 125 tumour-associated antigen: a review of the literature

    Human Reproduction

    (1989)
  • R.C Knapp et al.

    Clinical perspectives in using CA 125

    Contemporary Obstetrics and Gynecology

    (1996)
  • P.E Schwartz

    The role of tumor markers in the preoperative diagnosis of ovarian cysts

    Clinical Obstetrics and Gynecology

    (1993)
  • H.S Cuckle et al.

    Screening for ovarian cancer

  • N.J Finkler et al.

    Comparison of serum CA 125, clinical impression, and ultrasound in the preoperative evaluation of ovarian masses

    Obstetrics and Gynecology

    (1988)
  • E.M.J Schutter et al.

    Diagnostic value of pelvic examination, ultrasound, and serum CA 125 in postmenopausal women with a pelvic mass

    Cancer

    (1994)
  • I Jacobs et al.

    A risk of malignancy index incorporating CA 125, ultrasound and menopausal status for the accurate preoperative diagnosis of ovarian cancer

    British Journal of Obstetrics and Gynaecology

    (1990)
  • A.P Davies et al.

    The adnexal mass: benign or malignant? Evaluation of a risk of malignancy index

    British Journal of Obstetrics and Gynaecology

    (1993)
  • S Tingulstad et al.

    Evaluation of a risk of malignancy index based on serum CA 125, ultrasound findings and menopausal status in the pre-operative diagnosis of pelvic masses

    British Journal of Obstetrics and Gynaecology

    (1996)
  • Cited by (77)

    • Improving diagnostic strategies for ovarian cancer in Filipino women using ultrasound imaging and a multivariate index assay

      2022, Cancer Epidemiology
      Citation Excerpt :

      Since then, various studies have been conducted to assess their local diagnostic accuracy and applicability [4,6,7,24,25]. Subjective assessment by a level III expert sonologist [26] has been proven to be the most superior method in distinguishing benign from malignant ovarian masses[8,27–30], which was also the case for our center as IOTA-LR1 and LR2 were the best individual risk prediction classifiers. However, one major limitation of this method is its subjectivity and the requirement for an experienced examiner [31,32].

    • Methods of Assessing Ovarian Masses: International Ovarian Tumor Analysis Approach

      2019, Obstetrics and Gynecology Clinics of North America
      Citation Excerpt :

      Various diagnostic approaches have been introduced to characterize ovarian pathology before surgery. Most tests used in clinical practice incorporate findings of transvaginal ultrasound examination, because evidence shows this is the most appropriate first-line imaging technique for the preoperative assessment of women with adnexal pathology,8 with subjective evaluation of ultrasound findings by an experienced examiner being the best method for discriminating between benign and malignant disease.9–11 In a randomized controlled trial, the management of patients with adnexal lesions has been shown to benefit from ultrasound assessment by experienced operators.12

    • The Abnormal Ovary: Evolving Concepts in Diagnosis and Management

      2019, Obstetrics and Gynecology Clinics of North America
    • Classification systems and prediction of risk of malignancy of the adnexal masses

      2018, Clinica e Investigacion en Ginecologia y Obstetricia
    View all citing articles on Scopus
    View full text