CR-GLoCo: Cross-Resolution Learning via Global-Local Context Consistency for semi-supervised 3D medical segmentation

Abstract

Deep learning-based methods have achieved remarkable accuracy in segmenting three-dimensional (3D) medical images but they typically rely on large amounts of expert annotated datasets for training. Annotating such datasets, particularly at the voxel-level, is laborious and costly. To alleviate such annotation burden, semi-supervised learning (SSL) is introduced, which can leverage a small set of labeled data together with abundant unlabeled data for improved performance. Limited by GPU memory, patch- or slice-based strategies are often adopted to train such a SSL model, which may result in a loss of global anatomical context. Additionally, patch/slice-based losses often make the trained SSL models sensitive to contextual changes and hinder generalization under scarce labels. To this end, we propose CR-GLoCo, a cross-resolution learning framework that enforces Global-Local Context Consistency for semi-supervised 3D medical image segmentation. CR-GLoCo couples a low-resolution global branch with a high-resolution local branch and enforces prediction consistency on their overlapping field of view. In particular, the global branch provides holistic anatomical priors while the local branch preserves fine boundaries. We further design a mutual pseudo-supervision with confidence filtering to transfer reliable cues across resolutions and adopt an overlap-based sampling strategy to leverage changing contexts during training. We conduct comprehensive experiments on three typical yet challenging datasets to evaluate the performance of the proposed CR-GLoCo framework. Experimental results demonstrate that our proposed method achieves superior performance than other state-of-the-art SSL methods. Our code is available at https://github.com/luckieucas/CR-GLoCo.