Ovarian cancer classification with missing data

Unlike for other types of cancers, there are no tests suitable for mass screening that will reliably detect cases of ovarian cancer. The most commonly used test in clinical environments is the test for the tumour marker CA 125, with an accuracy of only up to 70 %. Research so far has mostly focused on predicting the risk of malignancy of pelvic masses, using operator-dependent ultrasound characteristics. We examine the use of standardized blood-test data for classification of benign cysts and malignant conditions. We use multi-layer perceptron (MLP) networks for classification and show that it is even possible to distinguish between early and late stage cancers with an accuracy of up to 92.9 %. Furthermore, we deal with missing data, an inevitable problem in clinical research that is not yet well explored in the context of neural networks. We compare several intuitive heuristics to estimate missing data and show that MLP networks are a suitable tool to deal with incomplete inputs.