Does a semi-supervised learning algorithm using conditional deep generative models improve left ventricle segmentation accuracy in echocardiography compared to standard algorithms?
A novel semi-supervised learning algorithm using generative models improves the accuracy of left ventricle segmentation in echocardiography by leveraging unlabeled frames.
Accurate segmentation of left ventricle (LV) in apical four chamber echocardiography cine is a key step in cardiac functionality assessment. Cardiologists roughly annotate two frames in the cardiac cycle, namely, the end-diastolic and end-systolic frames, as part of their clinical workflow, limiting the annotated data to less than 5% of the frames in the cardiac cycle. In this paper, we propose a semi-supervised learning algorithm to leverage the unlabeled data to improve the performance of LV segmentation algorithms. This approach is based on a generative model which learns an inverse mapping from segmentation masks to their corresponding echo frames. This generator is then used as a critic to assess and improve the LV segmentation mask generated by a given segmentation algorithm such as U-Net. This semi-supervised approach enforces a prior on the segmentation model based on the perceptual similarity of the generated frame with the original frame. This approach promotes utilization of the unlabeled samples, which, in turn, improves the segmentation accuracy.
Jafari et al. (Mon,) studied this question.