Vertebra segmentation from biplanar whole-spine radiographs is highly demanded in the quantitative assessment of scoliosis and the resultant sagittal deformities. However, automatic vertebra segmentation from the radiographs is extremely challenging due to the low contrast, blended boundaries, and superimposition of many layers, especially in the sagittal plane. To alleviate these problems, we propose a lightweight pyramid attention quick refinement network (LPAQR-Net) for efficient and accurate vertebra segmentation. The LPAQR-Net consists of three components: (1) a lightweight backbone network (LB-Net) to prune network parameters and memory footprints to strike an optimal balance between speed and accuracy, (2) a series of global attention refinement (GAR) modules to selectively reuse low-level features to facilitate the feature refinement, and (3) an attention-based atrous spatial pyramid pooling (A-ASPP) module to extract weighted pyramid contexts to improve the segmentation of blurred vertebrae. Moreover, the multi-class training strategy is employed to alleviate the over-segmentation of adjacent vertebrae. Evaluation results on both frontal and lateral radiographs of 332 AIS patients show our method achieves accurate vertebra segmentation with significant reductions in inference time and computational demands compared to the state-of-the-art. Meanwhile, results on the public AASCE2019 dataset also demonstrate the good generalization ability of our model. It is the first attempt to explore the lightweight network for vertebra segmentation from biplanar whole-spine radiographs. It simulates radiologists gathering nearby contexts for accurate and robust vertebra boundary inference. The method can provide efficient and accurate vertebra segmentation for clinicians to perform a fast and reproducible spinal deformity evaluation.