Keywords: Cardiac MRI Segmentation, Native T1 mapping, Uncertainty, Quality control
Abstract: Cardiac MR segmentation using deep learning has advanced significantly. However, inaccurate segmentation results can lead to flawed clinical decisions in downstream tasks. Hence, it is essential to identify failed segmentations through quality control (QC) methods before proceeding with further analysis. This study proposes a fully automatic uncertainty-based QC framework for T1 mapping and extracellular volume (ECV) analysis, consisting of three parts. Firstly, Bayesian Swin transformer-based U-Net was employed to segment cardiac structures from a native and post-contrast T1 mapping dataset (n=295). Secondly, our novel uncertainty-based QC, which utilizes image-level uncertainty features, was used to determine the quality of the segmentation outputs. It achieved a mean area under the ROC curve (AUC) of 0.927 on binary classification and a mean absolute error (MAE) of 0.021 on Dice score regression. The proposed QC significantly outperformed other state-of-the-art uncertainty-based QC methods, especially in predicting segmentation quality from poor-performing models, highlighting its robustness in detecting failed segmentations. Finally, T1 mapping and ECV values were automatically computed after the inaccurate segmentation results were rejected by the QC method, characterizing myocardial tissues of healthy and cardiac pathological cases. The myocardial T1 and ECV values computed from automatic and manual segmentations show an excellent agreement yielding Pearson coefficients of 0.990 and 0.975, respectively. The study results indicate that these automatically computed values can accurately characterize myocardial tissues.