search railabs.duhs.duke.edu.

Papers Published

1. Lo, J.Y. and Floys, C.E., Jr, Analysis of error surfaces of neural network applied to computer-aided diagnosis in mammography, WCNN'96. World Congress on Neural Networks. International Neural Network Society 1996 Annual Meeting (1996), pp. 1240 - .
   (last updated on 2007/04/15)

   Abstract:
   This study investigates the underlying behaviour of an artificial neural network (ANN) for computer-aided diagnosis in mammography. ANNs were previously developed that given patient age and only two mammographic features, accurately predicted biopsy outcome for 266 patients with breast lesions. The performance was comparable to other ANNs for the same purpose which required 14, 43, or more features. The simplicity of this three-input ANN suggested relatively well-behaved error surfaces in weight space. To analyze those error surfaces, the three-input ANN was simplified from a backpropagation network with one hidden layer to a single-node perceptron which still maintained diagnostic accuracy. Four-dimensional error surfaces were generated by independently varying the perceptron's three weights and bias value, while measuring performance by mean squared error and ROC area index. The error surfaces were finally visualized to reveal patterns of local and global minima

   Keywords:
   backpropagation;diagnostic radiography;error analysis;medical diagnostic computing;neural nets;