Abstract:    Ovarian cancer does not occur very frequently (only 4% of cancers among women) but is mostly diagnosed at a late stage because of the lack of symptoms at an early stage. Therefore ovarian cancer is often called "the silent killer" and clinical management of this disease is complicated by several factors. Firstly, it is difficult to distinguish benign and malignant ovarian masses before surgery based solely on clinical and ultrasound data. A correct diagnosis still largely depends on the experience of the clinician. This distinction is important because there is a favourable effect on the prognosis if the patient is directed immediately to an ovarian cancer specialist when diagnosed with an ovarian mass. Secondly, in the case of early stage disease it is difficult to foresee if the tumor will recur or not after surgery. An option in early stage disease is to give adjuvant therapy after surgery (e.g., chemotherapy) although it is known that some patients would not benefit from it (since they will never have a relapse without adjuvant therapy and are cured whatsoever) and they would only be subjected to unnecessary toxicity. At this moment however, no reliable clinical parameters are available that can predict recurrence in early stage disease. A third issue concerns treatment in advanced stage disease (FIGO stage III or IV). Some tumors will prove to be resistant to platin based chemotherapy (platin resistance). Correct prediction of platin resistance would enable medical doctors to supply patients suffering from this disease with realistic information about their prognosis and enable optimal management strategies.
The goal of this research project is to construct models that are able to discriminate benign from malignant masses before surgery, to predict recurrence after surgery in early stage disease and to predict the response to platin based chemotherapy in advanced stage disease. To achieve this, we will use Bayesian networks to model and combine different types of data: clinical, ultrasound, proteomic and microarray data. Proteomic data can result in a better pre-operative distinction between benign and malignant ovarian masses while both microarray and proteomic data can enhance the performance when predicting recurrence or the response to therapy.
Bayesian networks are a marriage between probability theory and graph theory. A Bayesian network is a sparse way of writing down a joint probability distribution. They are white box models which is important when encountering medical problems. When confronted with a certain question they give a complete overview of the uncertainty of the answer. Moreover such models allow exploring the underlying mechanism that generated the data and discover new (possibly causal) connections between the variables. Additionally Bayesian networks allow to incorporate prior knowledge (e.g. expert information, relevant literature) into the model. This allows guiding model training in the right direction. The usage of prior knowledge is very important with small and medium sample sizes that often occur in medical problems.