Deep learning has shown remarkable success in medical image analysis over the last decade; however, many contributions focused on supervised methods which learn exclusively from labeled training samples. Acquiring expert-level annotations in large quantities is time-consuming and costly, even more so in medical image segmentation, where annotations are required on a pixel level and often in 3D. As a result, available labeled training data and consequently performance is often limited. Frequently, however, additional unlabeled data are available and can be readily integrated into model training, paving the way for semi- or self-supervised learning (SSL). In this work, we investigate popular SSL strategies in more detail, namely Transformation Consistency, Student–Teacher and Pseudo-Labeling, as well as exhaustive combinations thereof. We comprehensively evaluate these methods on two 2D and 3D cardiac Magnetic Resonance datasets (ACDC, MMWHS) for which several different multi-compartment segmentation labels are available. To assess performance in limited dataset scenarios, different setups with a decreasing amount of patients in the labeled dataset are investigated. We identify cascaded Self-Supervision as the best methodology, where we propose to employ Pseudo-Labeling and a self-supervised cascaded Student–Teacher model simultaneously. Our evaluation shows that in all scenarios, all investigated SSL methods outperform the respective low-data supervised baseline as well as state-of-the-art self-supervised approaches. This is most prominent in the very-low-labeled data regime, where for our proposed method we demonstrate 10.17% and 6.72% improvement in Dice Similarity Coefficient (DSC) for ACDC and MMWHS, respectively, compared with the low-data supervised approach, as well as 2.47% and 7.64% DSC improvement, respectively, when compared with related work. Moreover, in most experiments, our proposed method is able to greatly decrease the performance gap when compared to the fully supervised scenario, where all available labeled samples are used. We conclude that it is always beneficial to incorporate unlabeled data in cardiac MRI segmentation whenever it is present.
Urschler et al. (Tue,) studied this question.