<p>Medical image segmentation plays an important role in supporting clinical diagnosis and treatment planning. Despite the success achieved by fully supervised learning methods, they heavily rely on costly manual annotations. Though semi-supervised learning is a promising approach to reducing the expensive annotation cost, existing methods have not fully exploited contextual information in both the spatial and (Fourier) spectral domains of medical images. To address this issue, we propose a Spatial-Spectral Mask Consistency learning method for semi-supervised medical image segmentation. In our method, unlabeled images are input into two parallel masking pathways of spatial and spectral domains for consistency learning. The first one randomly masks out a set of patches in the pixel space and constraints consistency between pixel-wise predictions of masked images and their originals. In this way, masked predictions are recovered based on visible pixels for contextual understanding. Since spatial masking may remove regions of interest (ROIs) inattentively, Fourier spectrum of an image is randomly masked out in the second pathway to preserve ROIs. Spectral masking serves as a data augmentation strategy to simulate possible medical image degradation, in which prediction consistency is constrained in both spatial and spectral domains for learning. Experiments demonstrate the superiority of our method over the state of the art for semi-supervised medical image segmentation.</p>