Exploring the Relationship Between Segmentation Uncertainty, Segmentation Performance and Inter-observer Variability with Probabilistic Networks

摘要

Medical image segmentation is an essential tool for clinical decision making and treatment planning. Automation of this process led to significant improvements in diagnostics and patient care, especially after recent breakthroughs that have been triggered by deep learning. However, when integrating automatic tools into patient care, it is crucial to understand their limitations and to have means to assess their confidence for individual cases. Aleatoric and epistemic uncertainties have been subject of recent research. Methods have been developed to calculate these quantities automatically during segmentation inference. However, it is still unclear how much human factors affect these metrics. Varying image quality and different levels of human annotator expertise are an integral part of aleatoric uncertainty. It is unknown how much this variability affects uncertainty in the final segmentation. Thus, in this work we explore potential links between deep network segmentation uncertainties with inter-observer variance and segmentation performance. We show how the area of disagreement between different ground-truth annotators can be developed into model confidence metrics and evaluate them on the LIDC-IDRI dataset, which contains multiple expert annotations for each subject. Our results indicate that a probabilistic 3D U-Net and a 3D U-Net using Monte-Carlo dropout during inference both show a similar correlation between our segmentation uncertainty metrics, segmentation performance and human expert variability.