Generalization Capacity of Neural Networks for the Classification of Ovarium Tumours

An experienced surgeant is able to make a prediction about the malignancy of an ovarium tumour by means of the presented medical information (such as medical images, morphology and demographic features). To support this classiication process, we try to automate the prediction by means of a supervised neural network. A neural network is able to deduce unknown relations from the data by means of its learning capacity and to generalise towards unseen cases (not presented in the training set). Therefore, it is suited for this biomedical classiication task. To obtain a good generalization, large amounts of data have to be presented to the neural network. Unfortunately, we only have a small data set at our disposal, such that the risk of memo-rising the examples in the training set increases. This phenomenon is called overtraining. Since a wrong diagnosis has very strong consequences on the welfare of the patient, the generalization capacity of the network has to be examined thoroughly. We will study two techniques to avoid overtraining: early stopping and regularization.