LGS: A Lightweight 4D Gaussian Splatting for Efficient Surgical Scene Reconstruction This paper introduces a lightweight 4D Gaussian Splatting (LGS) framework for efficient surgical scene reconstruction. The framework addresses the challenges of rendering efficiency and memory efficiency in dynamic surgical scenarios. LGS employs deformation-aware pruning to minimize the redundancy of Gaussian quantities, Gaussian-attribute pruning to reduce the dimensions of Gaussian attributes, and feature field condensation to tackle high-resolution redundancy in the spatial-temporal feature field. These techniques enable LGS to achieve a compression rate exceeding 9× while maintaining high-quality and real-time rendering. Experiments on public datasets demonstrate that LGS achieves a compression rate of 15× on ENDONERF and 9× on SCARED compared to EndoGaussian, and its rendering quality is comparable to GS-based methods. LGS also outperforms EndoGaussian on SCARED in SSIM and PSNR. The framework is optimized using knowledge distillation, which transfers knowledge from a well-trained model to a memory-efficient student model. LGS is shown to be as memory-efficient as NeRF-based methods while achieving high-quality and real-time rendering for dynamic surgical scenes. The results indicate that LGS is a significant step towards the practical deployment of 4D Gaussian Splatting in robotic surgical services.LGS: A Lightweight 4D Gaussian Splatting for Efficient Surgical Scene Reconstruction This paper introduces a lightweight 4D Gaussian Splatting (LGS) framework for efficient surgical scene reconstruction. The framework addresses the challenges of rendering efficiency and memory efficiency in dynamic surgical scenarios. LGS employs deformation-aware pruning to minimize the redundancy of Gaussian quantities, Gaussian-attribute pruning to reduce the dimensions of Gaussian attributes, and feature field condensation to tackle high-resolution redundancy in the spatial-temporal feature field. These techniques enable LGS to achieve a compression rate exceeding 9× while maintaining high-quality and real-time rendering. Experiments on public datasets demonstrate that LGS achieves a compression rate of 15× on ENDONERF and 9× on SCARED compared to EndoGaussian, and its rendering quality is comparable to GS-based methods. LGS also outperforms EndoGaussian on SCARED in SSIM and PSNR. The framework is optimized using knowledge distillation, which transfers knowledge from a well-trained model to a memory-efficient student model. LGS is shown to be as memory-efficient as NeRF-based methods while achieving high-quality and real-time rendering for dynamic surgical scenes. The results indicate that LGS is a significant step towards the practical deployment of 4D Gaussian Splatting in robotic surgical services.