Impact of segmentation uncertainties on computer-aided diagnosis of pulmonary nodules

Objective: Computer-aided diagnosis (CADx) systems can be utilized to estimate the likelihood of malignancy of pulmonary nodules detected on CT scans. Such systems often require an initial segmentation step. We report on the impact of segmentation variations on the CADx output and describe methods for reducing this effect. Methods: A total of 125 pulmonary nodules were evaluated using a CADx system based on genetic algorithms and ensemble classifiers. Two segmentation approaches were tested using leave-one-out validation: one approach included a free parameter chosen manually (manual) while the second used a parameter-free variant of the same algorithm (automatic). Results: By varying the free parameter for the manual segmentation of the testing data, a set of likelihoods of malignancy was computed for each nodule. The mean range of these likelihoods was 0.36 (±0.23) and 0.45 (±0.27) using the manual and automatic segmentation approaches for the training data, respectively. Using training and testing data segmented with the automatic approach yielded an Az of 0.87 (±0.03). Similar results were obtained when the training and testing data were segmented using either approach. Conclusion: Nodule segmentation can significantly affect CADx performance. A parameter-free segmentation was shown to match the performance of a skilled user while reducing potential user-to-user variability.