In recent years, artificial neural network (ANN) techniques have dominated the relatively new field of computer-aided diagnosis (CAD), particularly in the diagnosis of breast cancer. Most of these studies relied on a classic ANN paradigm, namely, the single-hidden-layer, fully-interconnected, feed-forward, error-back-propagation network, which used sigmoid activation functions. Although versatile and popular, this classic ANN approach has many limitations. In particular, the gradient descent technique used to train network weights is susceptible to entrapment in local minima. Furthermore, the number of hidden nodes are fixed and chosen arbitrarily. To achieve a desired level of performance, too many hidden nodes are frequently used, resulting in overfitting of the training cases, which compromises the network's ability to generalize to new cases it has not seen before.
To address these limitations, the evolutionary computing (EC) paradigm was investigated as an alternative to the classic ANN paradigm. The EC paradigm is a stochastic optimization technique, consisting of a blend of evolutionary programming (EP) and evolutionary strategies (ES), which numerically addresses (but is not immune from) the problem of entrapment in local minima.
Using available mammographic findings and patient history, researchers applied these techniques to the problem of predicting whether a breast lesion was benign or malignant. Mammographic findings were used because mammography is the most widely used radiologic modality for the early detection of breast cancer. However, only 15-34% of women who undergo a breast biopsy for a mammographically suspicious, nonpalpable lesion actually have breast cancer. Thus, 66-85% of the biopsies performed today could be avoided if these lesions could be classified accurately using the information from a mammogram.
Online access to SPIE eBooks is limited to subscribing institutions.