Semi-supervised Segmentation of Cardiac MRI using Image Registration
Refereed Conference Meeting Proceeding
In this work, we propose a novel method for enhancing the segmentation of Cardiac MRI through synthetically labelling unlabelled volumes and sequences within the dataset. In particular, the synthetic labels are obtained through registration using a Voxelmorph network trained over the labelled subset. Unlike other works that use 3D networks the segmentation is performed using a 2D U-net, resulting in a weakly supervised segmentation approach that benefits from the features in unlabelled volumes along with the annotated data. Experimental results for over 345 patients scanned using four different scanners show that the addition of the synthetically obtained labels to the original dataset enhance the performance up to 2.6% and improves the capability of the network to generalise to scanners from different vendors.
Digital Object Identifer (DOI):
Dublin City University (DCU)
Open access repository: