Recently, it has been shown that the uncertainty can be estimated by integrating Monte Carlo (MC) dropout [1] in the neural network structure in deep learning. Our group proposed Bayesian U-Net [2], which integrates MC dropout with U-Net [3] architecture, and investigated the relationship between the segmentation accuracy (Dice) and the uncertainty inferred by MC dropout. High correlation between the uncertainty and the errors in the automatic segmentation of musculoskeletal structures in CT images was obtained. In a different study, the integration of the segmentations obtained by voting based on prediction probabilities by convolutional neural networks (CNNs) trained for each anatomical plane independently [4] could further improve the segmentation accuracy and reduced the calculation time. However, as far as we know, the uncertainty-based integration has not been investigated yet. In this paper, we applied 3 Bayesian U-Nets, each trained on 2D slices in one of the three anatomical planes, to two-phase contrast-enhanced CT images of 48 cases (97 images) with manual segmentation of 17 organs. We report the variations in the uncertainties with respect to the anatomical planes, and finally evaluate the multiplanar integration-based predictions compared with single plane-based predictions. The segmentation accuracy, represented by Dice coefficient (DC), was significantly improved by the uncertainty-based integration in 9 organs (p<0.05) compared with segmentations obtained only from the axial plane. All segmentation results showed a negative correlation between the uncertainty and DC. We found that the segmentation accuracy could be improved by integrating the multiplanar segmentation results based on uncertainty estimation.
|