Background: Effective radiotherapy for thoraco-abdominal cancers relies on accurate 3D tracking of targets and adjacent organs at risk. However, the development and clinical implementation of motion-adaptive digital health solutions are hindered by limited access to high-quality, time-resolved 3D imaging with expert annotations. Open, reproducible 4D (3D+time) image datasets that capture multi-cycle respiratory variability and provide ground truth (GT) motion information would accelerate the translation of imaging research into real-time adaptive radiotherapy.

Objectives: This study aimed to generate and openly share anatomically plausible 4D image phantoms that (1) span multiple respiratory cycles, (2) provide perfect GT motion information, and (3) are built entirely from publicly available heterogeneous datasets. The resource is intended to support algorithm development, benchmarking, and implementation research, enabling reproducible and scalable evaluation of digital-health tools.

Methods: An imaging modality-agnostic pipeline was developed to integrate: (i) deformation vector fields (DVFs) derived from single-cycle 4D imaging datasets, (ii) static 3D target images with GT annotations, and (iii) multi-cycle respiratory traces. DVFs were aligned and resampled to the target image space, with spatial gaps completed via vector inpainting and distance-weighted attenuation. Temporal interpolation guided by respiratory traces generated multi-cycle DVFs, which were applied to warp static 3D targets and generate time-resolved 4D image sequences at an acquisition rate of five images per second.

Results: In total, sixty 4D phantoms were produced, including twenty CTs and forty magnetic resonance (MR) phantoms, spanning a range of anatomical configurations and motion patterns. All code, input data, and outputs are publicly released (https://github.com/laboratory-for-translational-medicine/4DPhantoms). Representative CT and MR phantoms with associated target and multi-organ segmentations are shown in Figure 1, highlighting both image realism and the availability of voxel-level GT annotations across modalities. The phantoms exhibited physiologically plausible volumetric changes and inter-cycle variability. Expert evaluation by two radiation oncologists, seven medical physicists, and eight biomedical scientists yielded a mean realism score of 4.2 ± 0.8 on a five-point scale. Comparison with real image datasets acquired on MR-guided linear accelerators during radiotherapy delivery further demonstrated that the synthetic phantoms enable reliable evaluation of motion-tracking algorithms under controlled yet clinically realistic conditions.

Conclusions: These open-access 4D phantoms provide a reproducible and scalable foundation for translational research. By combining realistic anatomy, multi-cycle respiratory motion, and perfect GT information, they support rigorous validation of motion-adaptive imaging technologies and can accelerate safe deployment of motion tracking and adaptive therapy tools, particularly in settings with limited access to advanced 4D imaging.

Figure 1