Ovarian cancer diagnosis with complementary learning fuzzy neural network
Artificial Intelligence in Medicine
| Hi-index | 0.00 |
Ovarian cancer has been the leading cause of death from gynecologic cancer in the United States. Seventy percent of women with ovarian cancer have advanced disease at diagnosis, which has resulted a low 5-year survival rate (\math30%) with no appreciable improvement in the past several decades, despite the increasing availability of treatment alternatives [6]. The dramatic difference in cure between patients with local disease (\math80%) and those with distant disease (15%-25%) [2] strongly indicates the need for a non-invasive, yet effective test that is applicable to a large group of at-risk population to detect ovarian cancer in its early stages.Currently the most common serum test for ovarian cancer is CA 125. Even though elevated level of CA 125 has been observed in 90% of patients with advanced ovarian cancer, its usefulness in large-scale screening effort is limited by a number of benign conditions that can also result in elevated CA 125 levels [4]. Among patients diagnosed with a pelvic mass, the specificity of CA 125 in distinguishing malignant from benign tumors decreases among pre-menopausal patients [3]. Since a family history of ovarian cancer is the most important risk factor for ovarian cancer and there have been evidence that inherited cancers are often diagnosed at earlier ages [5], the low specificity of CA 125 among pre-menopausal patients with pelvic masses hampers its use in evaluating such patients who are predisposed genetically to ovarian cancer. Woolas et al reported the combined use of multiple serum markers with several analytical techniques, including a) voting panels; b) logistic regression; and c) classification and regression tree (CART) analysis.The results from these approaches were superior to those of the individual assays in discriminating malignant from benign pelvic masses [7]. A limitation of the study was that the same data were used to derive the classification rules (selection of the optimal panels of markers and the estimation of regression function parameters) as well as to test the performance of the classification rules. In this paper, we introduce an ANN-based classification system to collectively analyze a panel of four selected serum tumor markers, CA 125 II, CA 72-4, CA 15-3, and lipid-associated sialic acid (LASA). The ANN classifier produces a single-valued diagnostic index to be used to discriminate malignant from benign pelvic masses. The performance of the ANN classifier is compared to those of the individual assays using independent test data. The ANN classifier reported in this paper for discriminating malignant from benign pelvic masses was constructed based on the multilayer perceptron (MLP) structure ([1], [8]); the most commonly used ANN in medicine. To compensate for the small training sample size and noisy data as often occurred in medical applications, special sample selection criteria are applied to improve data quality. Preprocessing steps based on biological knowledge and data, mining techniques are also taken to reduce the complexity of ANN training. The original data set was divided into two sets, one for ANN training set and the other for independent validation. Two additional independent data sets were also used for the evaluation of the system.