Uncertainty-based method for improving poorly labeled segmentation datasets