Quantitative analysis of lymph node volume is instrumental in the diagnosis and treatment of cancer. However, automatic segmentation models for lymph nodes necessitate pixel-level labeling, which is both time-consuming and labor-intensive. The scarcity of pixel-level annotations has thus spurred interest in label-efficient learning as a potential solution. Considering the variance of shapes and locations, and the low-contrast appearance of lymph nodes in computed tomography scans, we propose a new incomplete annotation strategy called orthogonal partial-instance annotation, in which only two orthogonal slices of a small portion of lymph nodes are annotated. To segment as many lymph nodes as possible from such sparse annotations, we propose a prototype-based label-efficient learning framework with a specifically designed loss. Specifically, we extract intra-batch prototypes from the output features of the encoder and store inter-batch prototypes using a momentum-smoothing approach. To re-inject the extracted information from the two kinds of prototypes, we introduce a feature augmentation module that utilizes the extracted prototypes to enhance features. To further complement the predictions generated from enhanced features with those from original features, we design a reliability-based co-teaching strategy based on feature similarity. Experiments demonstrate that our proposed framework outperforms other methods on two mediastinal lymph node datasets. Our implementation is available at https://github.com/HiLab-git/WCODE-PIA .