The semi-supervised method has greatly promoted the development of medical image segmentation, as it alleviates the pressure of obtaining a large number of medical image annotations. In this article, we propose a new spatial-spectral and commonality learning network (SSCL) that better utilizes unlabeled data for semi-supervised medical image segmentation. The motivation of the SSCL model is to observe that, existing methods only focus on the features of the target area during segmentation, ignoring the features which can represent the total image, we call them common features. Because there is also feature information in the common features that can better assist the network in segmentation. In addition, due to the low contrast and high noise characteristics of medical images, only allowing the model to learn features in the spatial domain is not sufficient for the network to learn enough information. Due to insufficient feature information, the network will make more erroneous predictions when segmenting the edges of the target area than when segmenting the central area. Therefore, our proposed SSCL model consists of two new designs to address the above issues. First, we propose a reliable commonality learning module to learn the common features to help the network improve the segmentation performance. Second, we design a spectral convolution module to learn spectral feature information. Experimental results on three medical image datasets show that our framework outperforms previous state-of-the-art methods.