Combining Boundaries and Ratings from Multiple Observers for Predicting Lung Nodule Characteristics

We use the data collected by the Lung Image Database Consortium (LIDC) for modeling the radiologists' nodule interpretations based on image content of the nodule by using decision trees. Up to 4 radiologists delineated nodule boundaries and provided ratings for nine nodule characteristics (lobulation, margin, sphericity, etc). Therefore, there can be up to 4 instances per nodule in our data set. However, to learn a good predictive model, the data set should have only one instance per nodule. In this study, we investigate several approaches to combine delineated boundaries and ratings from multiple observers. From our experimental results, we learned that the thresholded p-map analysis approach with the probability threshold PrGt=0.75 provides the best predictive accuracies for the nodule characteristics. In the long run, we expect that the predictive model will improve radiologists' efficiency and reduce inter-reader variability.