To alleviate the inefficiency of the current visualization methods in skipping the empty voxels located in the hollow part, this paper presents a GPU-based ray casting method for hollow volume data. First, the volume is divided into two equal parts and reconstructed into two octrees before rendering. In this process, the empty blocks are culled synchronously from the original data, and two vertex arrays are generated consisting of valid voxels only. Then, the vertex arrays are rendered into textures which contain the start and direction information of the ray. The final images are obtained by calculating the color and opacity of the two parts with CUDA, respectively. The experimental results show that our method can speed up the rendering speed considerably on the volume data with large hollow structure without any quality loss compared with the traditional visualization methods.