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

Abstract

The scarcity of labeled data often impedes the application of deep learning to the segmentation of medical images. Semi-supervised learning seeks to overcome this limitation by leveraging unlabeled examples in the learning process. In this paper, we present a novel semi-supervised segmentation method that leverages mutual information (MI) on categorical distributions to achieve both global representation invariance and local smoothness. In this method, we maximize the MI for intermediate feature embeddings that are taken from both the encoder and decoder of a segmentation network. We first propose a global MI loss constraining the encoder to learn an image representation that is invariant to geometric transformations. Instead of resorting to computationally-expensive techniques for estimating the MI on continuous feature embeddings, we use projection heads to map them to a discrete cluster assignment where MI can be computed efficiently. Our method also includes a local MI loss to promote spatial consistency in the feature maps of the decoder and provide a smoother segmentation. Since mutual information does not require a strict ordering of clusters in two different assignments, we incorporate a final consistency regularization loss on the output which helps align the cluster labels throughout the network. We evaluate the method on three challenging publicly-available datasets for medical image segmentation. Experimental results show our method to outperform recently-proposed approaches for semi-supervised segmentation and provide an accuracy near to full supervision while training with very few annotated images

An Embarrassingly Simple Consistency Regularization Method for Semi-Supervised Medical Image Segmentation

“…As seen in the table, our approach uses a low number of labelled samples and produces the best results in terms of DSC and HD, whereas methods like [2] use double the amount of labelled data, yet achieve inferior performance as compared to our results, proving the efficiency and robustness of our model. Only Peng et al [7] uses a lower number of labelled data, but produces poorer results as well. Our method produces a reasonably better result by making use of only 10% of the labelled data available for training.…”

“…All the other parameters were the same as those methods, to maintain fair comparison. Peng et al [7] use the least number of labelled samples in their method, but they have produced vastly substandard results. Whereas, our method obtains better results as compared to most state-of-the-art methods, by using 40% labelled data in the entire training set.…”

Local contrastive loss with pseudo-label based self-training for semi-supervised medical image segmentation

“…(c) Semi-supervised local contrastive learning: Recently, some concurrent works [25], [26], [47], [48], [49], [50], [51] have devised semi-supervised learning frameworks that use unlabeled images with some variant of contrastive loss setup. The works relevant to presented work are [25], [26].…”

Inconsistency-aware Uncertainty Estimation for Semi-supervised Medical Image Segmentation

“…Reg. [44], MSE [44], KL [44] and Peng et al 's method [23]. We also list the fully-supervised setting as a reference for the upper bound of semi-supervised segmentation methods.…”

“…Since all these methods are evaluated under the same setting, we directly report their performance presented in their literature. Note that, since the number of slices along z-axis is sometimes limited (i.e., less than 10), thus making 3D convolutions hard to perform, above methods usually adopted the 2D setting [53] [44] [23]. For fair comparison, we follow their 2D setting.…”