Improvement of detection accuracy in digital mammography with a pruned neural net optimized from heuristic decision rules

A pruned neural network approach is developed to improve detection accuracy of microcalcification in digital mammography. The architecture (both number of neurons and connections) of this neural network is designed based on the node operations and conditioning statements in a binary decision tree derived from a heuristic decision rule. The initial weights of the pruned neural net are initially configured with the same heuristic rule, then further optimized with backpropagation training. Preliminary results show a significant improvement in detection accuracy over winner-take-all strategy and original heuristic decision rule