3D Gaussian splatting can produce cinematic renderings of CT volumes, but real-time displays limit how many Gaussians can be used. In current practice, each patient's scene is optimized individually, and densification decides where the Gaussians go. The authors of this paper propose a reusable, label-aware allocation front end that instead decides where a fixed budget of B Gaussians should be placed, leaving the other attributes to be fitted per patient.

Given a patient's CT, anatomical labels, and transfer function, the front end produces a density over 2 mm cells and draws exactly B positions from it. It is trained once on per-patient optimized scenes and comes in two forms: a population-trained neural prior and an interpretable seven-parameter visibility-aware closed form (VACF).

On 20 held-out patients at budgets from 10K to 100K Gaussians, both methods raised vessel-masked PSNR by 4.5–7.1 dB over uniform allocation and by 1.6–3.3 dB over pruning a dense voxel-aligned model. The neural prior did not differ significantly from per-patient optimized allocation at any budget, and VACF was within 0.6 dB of it on vessels. On 19 external patients, the neural prior again matched per-patient optimization, with VACF within 0.7 dB.

Used as an initialization for position optimization, VACF beat per-patient densification on vessels at 10K Gaussians and, at 25K, reached a higher vessel PSNR after 118 seconds of optimization than densification did after 248 seconds. The two variants also differ in transfer: VACF generalizes to new colors and lighting but not to a label-free transfer function, while the neural prior shows the opposite pattern.