Decision-making differences: novices, experts, and a neural network

We investigated the decision making performance of trained radiographers, novice radiographers and a neural network in the detection of fractures. Ground truth was established by the independent agreement of experienced radiologists for 740 single view digitized radiographs of the wrist. The images were categorized into negatives and positives; 520 of these were used to train the back propagation, three layer neural network in a supervised mode, and the remainder were used to create a test bank. The test was presented to 20 novice observers, 12 experienced radiographers trained in the detection of skeletal trauma and then to the trained neural network. ROC Az values for all the decision makers were not significantly different (p > 0.1) but there were significant differences in the values of True Positive and True Negative Fractions. The neural network showed a greater aptitude for distinguishing the normals. By filtering the neural net decisions through the human data we simulated the effect of assisted reporting. Results suggest that if fracture prevalence is very low in a population, a neural network demonstrating high specificity may have utility in reducing the number of images which must be reviewed by human experts.