(Sunday, 7/14/2019) 3:00 PM - 3:30 PM
Room: Exhibit Hall | Forum 6
Purpose: With typically only one ground truth segmentation per image, many modern neural networks are unable to produce an uncertainty estimation. In this study, we utilize an efficient Bayesian approximation presented by Gal and Ghahramani, and apply it to quantify segmentation uncertainty on prostate tumor.
Methods: We used dropout and the Monte Carlo (MC) estimations during the training the prediction phase of the model, respectively. Utilizing 127 training, 42 validation, and 26 test prostate cancer patients, we localized the prostate in an 96 x 96 x 48 array, with voxel size of 1.17mm x 1.17mm x 2mm. We trained a U-net with ResNeXt blocks to learn the prostate segmentation. To evaluate the uncertainty, we developed a score function that rewards high uncertainty in mislabeled voxels and low uncertainty in correctly labeled voxels, and penalizes the opposite. The uncertainty score is designed to range from -1 (worst) to 1 (best). To avoid bias, the score is only calculated in a set of voxels of interest, voi, defined as 1% of the maximum uncertainty.
Results: The segmentation predictions acheived a high average Dice coefficient of 0.87 on the test data. The uncertainty, obtained using 20 MC estimations per patient, highly correlates to the errors between the ground truth and prediction. On average we achieved a positive uncertainty score of 0.0979 ± 0.0407, indicating that, the high and low uncertainties are correctly aligned to the mislabeled and correctly-labeled voxels. There is one patient that had a negative uncertainty score of -0.0178, but the magnitude is much smaller than than the mean score.
Conclusion: The prediction will contain the contour along with the uncertainty, which can be designed to alert the user/developer to be cautious about high uncertainty regions. This is particularly important in the medical field, where patient safety takes priority.
Funding Support, Disclosures, and Conflict of Interest: This study is supported by Cancer Prevention and Research Institute of Texas (CPRIT) (IIRA RP150485, MIRA RP160661)
Keywords
Taxonomy
Contact Email
