Anatomica：面向解剖扩散模型的几何与拓扑性质局部化控制
Anatomica: Localized Control over Geometric and Topological Properties for Anatomical Diffusion Models
提供一种推理时、基于长方体控制域的可微惩罚引导机制，可把几何（矩）与拓扑（持续同调）约束组合施加到扩散采样过程；对做可控生成、结构约束采样或医学形状合成的研究者，这是一个可直接迁移到潜扩散模型上的约束注入范式。
原文依据 we use cuboidal control domains of varying dimensionality, location, and extent to isolate relevant substructures and compute differentiable penalty functions
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论解剖结构的可控生成使得面向虚拟仿真试验与机器学习流程的合成数据集能够被理性设计。我们提出一种推理时引导框架，用于生成具有局部化几何与拓扑控制的 3D 多类解剖分割。在推理过程中，我们使用维度、位置与范围各异的立方体控制域来隔离相关子结构，并计算可微惩罚函数，将样本引导至目标约束。我们通过体素矩来施加尺寸、形状、位置与朝向等几何特征约束，而连通分量、环与空洞等拓扑特征则通过持续同调加以约束。最后，我们将该框架适配到潜扩散模型，其中神经场解码器可以部分地提取子结构，从而实现对解剖特征的高效测量与控制。该形式化方法开启了一个丰富的设计空间，其中多个约束可以被组合起来，以控制定义在任意维度与坐标系上的复杂结构。我们在 https://github.com/kkadry/Anatomica 发布代码。
英文原文摘要 1183 字符
Controllable generation of anatomical structures enables the rational design of synthetic datasets for virtual simulation trials and machine learning workflows. We present an inference-time guidance framework for generating 3D multi-class anatomical segmentations with localized geometric and topological control. During inference, we use cuboidal control domains of varying dimensionality, location, and extent to isolate relevant substructures and compute differentiable penalty functions that steer samples toward target constraints. We enforce geometric features such as size, shape, position, and orientation via voxel-wise moments, while topological features such as connected components, loops, and voids are enforced through persistent homology. Finally, we adapt this framework for latent diffusion models, where a neural field decoder can partially extract substructures, enabling efficient measurement and control of anatomical features. This formulation unlocks a rich design space where several constraints can be composed to control complex structures defined over arbitrary dimensions and coordinate systems. We release our code at https://github.com/kkadry/Anatomica.