Meta-learning for Medical Image Segmentation Uncertainty Quantification

Inter-rater and intra-rater variability is a major challenge in medical image segmentation. Inconsistencies of manual segmentations between different experts can challenge development of deterministic automated medical image analysis tools. QUBIQ 2021 is a challenge to enable the successful development of automated machine learning tools, when there are inconsistencies between the labels of different annotators. In this paper, we propose to use meta-learning for quantifying uncertainty in biomedical image quantification. We first train a segmentation network for each expert separately with extensive data augmentation using the nnUnet framework. Then, a meta learner model based on a conventional U-net architecture is trained using the average of all annotators as ground truth, and output of all models that have been trained for each radiologist as input. We compared our results of meta-learning with ensemble methods for various image segmentation tasks and illustrate improved performance.