LINGO: Latent Initialization and Gradient Optimization for Sparse-view X-ray Novel View Synthesis and CT Reconstruction with 3D Gaussian Splatting

arXiv cs.GR details LINGO, a unified 3D Gaussian Splatting framework aimed at sparse-view X-ray novel view synthesis and CT reconstruction. The pitch is straightforward: when angular coverage is thin, reconstruction gets noisy and structurally ambiguous, so the model tries to give the optimizer better priors up front and stronger gradients where the data is weakest.

LINGO combines latent mask-space initialization with dynamic gradient optimization. In practice, that means voxel-level 3D filters derived from X-ray masks to suppress background noise, plus adaptive voxel scaling and dynamically scaled loss terms to help low-density structures contribute more during training. The authors also introduce IPSD, a metric for measuring initialization quality.

For developers working on reconstruction, rendering, or other geometry-heavy ML systems, the practical takeaway is speed as much as quality. On the X3D dataset, the method reportedly reaches reconstruction quality comparable to state-of-the-art models in about 5k steps, versus roughly 30k iterations for the stronger baselines, while also improving PSNR and SSIM in both novel view synthesis and CT reconstruction.

The broader context is that 3DGS keeps expanding beyond real-time graphics into medical and scientific imaging, where initialization and gradient flow can matter as much as model capacity. If the gains hold up outside the benchmark setup, the approach could be useful anywhere sparse observations make optimization unstable.