We introduced the first open-sourced dataset for X-ray-guided miniature medical devices (MMD) [1], featuring stent-structured soft MMD [2,3] and shape-morphing ferrofluid MMD [4].

This dataset is organized into two domains, real and synthetic domain. The real domain comprises of 2 subsets (D1 and D2), with the former showcasing real-time MMD locomotion across diverse physiological environments and the latter capturing MMDs under varying imaging conditions. Synthetic dataset (D3) is generated by programmatically integrating MMD images into diffusion-model-generated tissue backgrounds. By making this dataset publicly available, we provide a foundational resource to train and benchmark MMD detection, segmentation, and tracking algorithms under X-ray imaging. This initiative facilitates comparative studies of deep learning architectures and training techniques, fostering reproducible research and community-driven progress in medical AI.

Label formatting followed a one-text-file-per-image protocol, with each text file sharing the corresponding image’s filename and a ‘.txt’ extension. Each row in the text file corresponded to a single object instance, annotated with class index and polygonal contour coordinates in the format: class index, x1, y1, x2, y2 ... xn, yn.

[1] Wang, Chunxiang, et al. "Synthetic X-Ray Driven Control of Miniature Medical Devices." 2026.

[2] Wang, Chunxiang, et al. "Heterogeneous multiple soft millirobots in three-dimensional lumens." Science Advances 10.45 (2024): eadq1951.

[3] Wang, Tianlu, et al. "Adaptive wireless millirobotic locomotion into distal vasculature." Nature communications 13.1 (2022): 4465.

[4] Sun, Mengmeng, et al. "Exploiting ferrofluidic wetting for miniature soft machines." Nature Communications 13.1 (2022): 7919.

Downloads last month: 2,286