Acquiring multiple images of the same patient (e.g., mediolateral oblique and craniocaudal view mammograms) can, in principle, help improve diagnostic accuracy. We investigated theoretically, in the context of computer-aided diagnosis (CAD), four methods of combining multiple computer outputs obtained from multiple images of the same patient: taking the average, the median, the maximum, or the minimum of the individual assessments. We assumed that multiple computer outputs for each patient are equally accurate and that they can be transformed monotonically to the same pair of truth conditional normal distributions. We found that both the average and the median always produce improved area under the ROC curve (AUC) compared to single-view images, and that the average always performs better than the median. Furthermore, the maximum and the minimum can also produce improved AUCs and can outperform the average under certain situations, but in other situations they can produce worse results than single-view images. Moreover, except for the median, each method can be the best-performing method under specific conditions. Finally, as the strength of correlation between image pairs increases, the maximum and the minimum tend to perform the best more often whereas the average is less often the best performer.
We have shown previously that there is uncertainty associated with the output of artificial neural network (ANN) and we have now developed a new method to reduce this uncertainty by training ANNs with multiple target values. In conventional ANN training, binary target values are used to represent, e.g., benign and malignant cases. However, this method does not take into consideration the various histology subtypes. In this work, we used both simulated datasets and a mammography dataset to show that the conventional training method leads to larger uncertainty in the ANN output. Eight ANNs were trained by choosing different initial weights and ANN output variance was measured by the average standard deviation (SD) of the 8 ANNs' outputs for each test case. In the simulation, in addition to the conventional training method using binary target values, we also trained ANNs with multiple target values, and a set of continuous target values derived from a likelihood ratio of the underlying distributions. For the mammogram study, we assigned multiple target values based on histology subtypes. Both the simulation and mammogram studies showed that ANNs produce very close overall performance regardless the training methods. However, training neural networks with multiple target values demonstrated lower uncertainty in the ANN outputs.