Diffusion Model-Based Hierarchical Registration of Whole-Body PET/CT Images

Whole-body PET/CT imaging technology can capture entire human-body images in a single scan to provide comprehensive metabolic and anatomical information. By analyzing these whole-body images, physiological and pathological changes in the whole body can be better understood. Since images can be acquired from different time points of the same subject (or different subjects), it is important to register them together for accurate measurement of signal changes in the same location. However, current registration methods are often designed for specific organs or regions, and thus cannot be effectively applied to whole-body image registration. To address this issue, we design a hierarchical registration framework based on diffusion model for whole-body images. The key components of our proposed framework include 1) designing a multi-part partitioning and fusion strategy to divide the whole-body image into multiple parts for reducing difficulty and computational burden of registration, 2) employing a diffusion model based deformable registration network to achieve precise registration for each corresponding part, and 3) further incorporating various topology constraints to preserve topological details for achieving more reliable registration. Extensive evaluations on 64 samples of whole-body images demonstrate that our proposed hierarchical framework can obtain robust registration across all human body parts, leading to accurate whole-body registration.