Boosting Semi-supervised Image Segmentation with Global and Local Mutual Information Regularization

The scarcity of labeled data often impedes the application of deep learning\nto the segmentation of medical images. Semi-supervised learning seeks to\novercome this limitation by exploiting unlabeled examples in the learning\nprocess. In this paper, we present a novel semi-supervised segmentation method\nthat leverages mutual information (MI) on categorical distributions to achieve\nboth global representation invariance and local smoothness. In this method, we\nmaximize the MI for intermediate feature embeddings that are taken from both\nthe encoder and decoder of a segmentation network. We first propose a global MI\nloss constraining the encoder to learn an image representation that is\ninvariant to geometric transformations. Instead of resorting to\ncomputationally-expensive techniques for estimating the MI on continuous\nfeature embeddings, we use projection heads to map them to a discrete cluster\nassignment where MI can be computed efficiently. Our method also includes a\nlocal MI loss to promote spatial consistency in the feature maps of the decoder\nand provide a smoother segmentation. Since mutual information does not require\na strict ordering of clusters in two different assignments, we incorporate a\nfinal consistency regularization loss on the output which helps align the\ncluster labels throughout the network. We evaluate the method on four\nchallenging publicly-available datasets for medical image segmentation.\nExperimental results show our method to outperform recently-proposed approaches\nfor semi-supervised segmentation and provide an accuracy near to full\nsupervision while training with very few annotated images.\n