Specialized 3D Gaussian Splatting Reconstruction Using Inner Gaussian Splatting and Drop Gaussian Techniques

Three-dimensional (3D) reconstruction technology plays a crucial role in the field of MRI image diagnosis. 3D reconstruction technology allows for the intuitive understanding of complex anatomical structures. Recently, 3D Gaussian Splatting (3DGS) has gained attention for its high rendering quality and speed, but it is primarily specialized in the surface reconstruction of objects. This makes it difficult to apply directly to MRI data, where internal structural information is vital. Furthermore, the slice-based acquisition of MRI data poses a sparse view problem from the perspective of 3D reconstruction, leading to overfitting and artifacts during 3DGS training. In this paper, we propose a 3DGS reconstruction method specifically tailored for MRI data to address these issues. The proposed method combines InnerGS technique for internal structure representation and DropGaussian technique to prevent overfitting in sparse views. Experimental results using the ADNI dataset show that the proposed method effectively reconstructs internal structures and demonstrates significant performance in both quantitative and qualitative evaluations.