FactorSplat: Appearance-Controllable Gaussian Proxies for Medical Volume Rendering

FactorSplat：用于医学体渲染的外观可控高斯代理

Transfer functions (TFs) control color and visibility in medical volume rendering, but image-trained Gaussian proxies typically bake one transfer function into their appearance. We present FactorSplat, a per-scene N-dimensional Gaussian splatting (N-DGS) proxy that accepts region-specific intensity-to-RGBA curves at inference.

传递函数（TF）在医学体渲染中控制颜色和可见性，但基于图像训练的高斯代理通常将单一的传递函数固化在其外观中。我们提出了 FactorSplat，这是一种针对场景的 N 维高斯溅射（N-DGS）代理，它支持在推理阶段输入特定区域的强度到 RGBA 的映射曲线。

A local lookup applies the authored color and opacity change, while a shared functional encoder and low-rank per-Gaussian factors learn the residual appearance response. Geometry and directional appearance remain shared across presets, with visibility control and TF-aware pruning preserving the ability to hide and reveal structures.

通过局部查找表应用预设的颜色和不透明度变化，同时利用共享的功能编码器和每个高斯点对应的低秩因子来学习残差外观响应。几何结构和方向性外观在不同预设间保持共享，通过可见性控制和感知传递函数的剪枝技术，保留了隐藏和显示特定结构的能力。

On seven CT and MR scans, FactorSplat improves mean PSNR and changed-region error over region-aware VEG across validation, interpolation, unseen composition, and out-of-distribution (OOD) edits. Across these four splits, seven-scan mean PSNR gains over VEG range from 1.10 to 1.52 dB. One checkpoint per scan supports unseen edits without retraining. At $1600^2$, the cached fast renderer averages 524 FPS with 1.17 ms TF switches.

在七项 CT 和 MR 扫描数据上，FactorSplat 在验证集、插值、未见过的组合以及分布外（OOD）编辑任务中，其平均峰值信噪比（PSNR）和变化区域误差均优于区域感知模型 VEG。在这四种测试划分中，七项扫描的平均 PSNR 相比 VEG 提升了 1.10 到 1.52 dB。每个扫描仅需一个检查点即可支持未见过的编辑，无需重新训练。在 $1600^2$ 分辨率下，缓存的快速渲染器平均达到 524 FPS，且传递函数切换仅需 1.17 毫秒。