Abstract
Electrocardiogram (ECG) delineation, the segmentation of meaningful waveformfeatures, is critical for clinical diagnosis. Despite recent advances usingdeep learning, progress has been limited by the scarcity of publicly availableannotated datasets. Semi-supervised learning presents a promising solution byleveraging abundant unlabeled ECG data. In this study, we present the firstsystematic benchmark for semi-supervised semantic segmentation (SemiSeg) in ECGdelineation. We curated and unified multiple public datasets, includingpreviously underused sources, to support robust and diverse evaluation. Weadopted five representative SemiSeg algorithms from computer vision,implemented them on two different architectures: the convolutional network andthe transformer, and evaluated them in two different settings: in-domain andcross-domain. Additionally, we propose ECG-specific training configurations andaugmentation strategies and introduce a standardized evaluation framework. Ourresults show that the transformer outperforms the convolutional network insemi-supervised ECG delineation. We anticipate that our benchmark will serve asa foundation for advancing semi-supervised ECG delineation methods and willfacilitate further research in this domain.