Semi-supervised 3D Myocardial Segmentation for Whole-Heart Joint T1/T2 mapping with Self-trained nnUNet

Motivation: Segmenting 3D whole-heart T1/T2 mapping data is time-intensive and automatic solutions require significant manual effort for labeling. A semi-supervised approach using unlabeled data could reduce this workload, accelerate segmentation, and enhance efficiency. Goal(s): To automate 3D myocardial segmentation for whole-heart joint T1/T2 mapping using nnUNet in a semi-supervised manner, using limited labeled data. Approach: A 3D nnUNet with pseudo-labeling is implemented for myocardial segmentation in joint T1/T2 mapping, combining manual labels and generated pseudo-labels to optimize the process. Results: This semi-supervised approach achieved a Dice above 0.877 using only 10 manual labels, with 80% fewer labels than supervised methods. Impact: Semi-supervised 3D nnUNet enables accurate myocardial segmentation in 3D whole-heart joint T1/T2 mapping, even with limited labeled data. This could improve efficiency, reduce manual segmentation effort, and accelerate the diagnosis of myocardial diseases.