Application of Fuzzy-Based Uncertainty in Cardiac MRI Segmentation

Cardiac magnetic resonance imaging (MRI) is pivotal in diagnosing cardiac-related diseases. Tissue segmentation from cardiac MRI images is the initial and most crucial step in downstream analyses. However, most current research focuses on designing segmentation models on cardiac MRI to produce segmentation results, without considering the reliability of these outputs. This paper novelty applied fuzzy-based uncertainty to assess the reliability of cardiac MRI segmentation results without access to ground truth images. Experimental results show that class-level uncertainty has a strong linear negative relationship (PE=-0.92) with the true segmentation quality (measured by Dice). This suggests that when ground truth images are unavailable, uncertainty can effectively estimate the cardiac MRI segmentation quality. Furthermore, qualitative analysis with clinicians is conducted to explore the clinical application of fuzzy-based uncertainty. Based on their feedback and comments, class-level uncertainty provides more detailed information and is more suitable to infer the segmentation quality in comparison to other uncertainties. As for the representation of uncertainty, the predicted Dice is the best option compared to other representation methods and is straightforward for clinicians to understand.