Investigation of regularized neural networks for the computerized detection of mass lesions in digital mammograms

The effectiveness of regularization is evaluated as a technique to minimize over-training in artificial neural networks and adds an extra term to the cost-function used in neural network training that penalizes over-complex results.

Abstract

Computerized schemes are currently being developed at the University of Chicago to detect mass lesions in digital mammograms. Artificial neural networks play an important role in the detection of masses. Currently, features are extracted from potential lesion areas and sent through a neural network to decide whether the area is to be called a true lesion or a false detection. One of the most difficult aspects of dealing with artificial neural networks is to train them without over-training; in other words, to take both the bias and variance into account when training. Typically, an early stopping technique is employed; that is, the neural network is tested on an independent data set and training is stopped when the the performance on this independent data set is maximized. In this paper the effectiveness of regularization is evaluated as a technique to minimize over-training. Regularization adds an extra term to the cost-function used in neural network training that penalizes over-complex results. The results of simulation studies will be presented along with results obtained using data of actual lesions and false positives from the authors' computerized mass detection scheme.