The scarcity of labeled data often impedes the application of deep learning the segmentation of medical images. Semi-supervised learning seeks to this limitation by exploiting unlabeled examples in the learning. In this paper, we present a novel semi-supervised segmentation method leverages mutual information (MI) on categorical distributions to achieve global representation invariance and local smoothness. In this method, we the MI for intermediate feature embeddings that are taken from both encoder and decoder of a segmentation network. We first propose a global MI constraining the encoder to learn an image representation that is to geometric transformations. Instead of resorting to-expensive techniques for estimating the MI on continuous embeddings, we use projection heads to map them to a discrete cluster where MI can be computed efficiently. Our method also includes a MI loss to promote spatial consistency in the feature maps of the decoder provide a smoother segmentation. Since mutual information does not require strict ordering of clusters in two different assignments, we incorporate a consistency regularization loss on the output which helps align the labels throughout the network. We evaluate the method on four publicly-available datasets for medical image segmentation. results show our method to outperform recently-proposed approaches semi-supervised segmentation and provide an accuracy near to full while training with very few annotated images.
No takes yet. Share an insight, caveat, or question.
Peng et al. (2021) studied this question.