Med Phys. 2026 Nov;53(11):e70742. doi: 10.1002/mp.70742.
ABSTRACT
BACKGROUND: In the growing field of interventional radiology, emerging technologies implement the use of artificial intelligence (AI) to help tackle growing demands for treatment. Due to the anatomical differences in animals, volunteers offer the most realistic training and testing scenario, however safety considerations restrict rapid prototyping of these new technologies. Simulators have been a key facilitator in the testing of new technologies among numerous medical engineering studies, but often lack the realism to reliably test AI.
PURPOSE: This proof-of-concept study presents methods to create or enhance hyper-realistic radiographic simulators that can effectively function with a medical-grade AI software.
METHODS: Iron-filled plastic was used to three-dimensionally (3D) print a series of blocks of heights up to 30 mm. By analyzing the intensities generated by different thicknesses of plastic, anonymous fluoroscopic images were turned into height maps (lithophanes) and 3D-printed. A static scene from transcatheter aortic valve implantation (TAVI) was then 3D-printed and imaged with different presets using two different x-ray systems for mathematical analysis. Lithophanes were then used as a radiographic background to experimental and commercial interventional simulators to test Caranx Medical's TAVIPILOT AI. Finally, fluoroscopies from 5 3D-printed lithophanes were mixed with 5 real images and presented to 5 expert volunteers to establish their realism in a structured survey.
RESULTS: The 3D-printed x-ray lithophanes test-plates' radiographic appearance was rated as "Adequate" to "Good" across all fields of the survey, with participants praising their realism. The test-plate was able to reliably recreate the TAVI scene in different x-ray systems, with joint-histogram correlation coefficients above 0.9, and global structural similarity above 0.8. The lithophanes were imaged as a standalone test plate and as a backing to dynamic TAVI interventional simulators; in both cases, TAVIPILOT AI responded to the simulator as if it was operating on a real anatomical image.
CONCLUSIONS: Lithophane plates present an easy-to-implement solution to recreate monoplane x-ray scenes, which can be deployed to test software at the point of care with minimal logistical, ethical, or safety concerns. This study opens the door for a new range of interventional simulators able to recreate human-like appearance to test AI, offering a non-exclusive alternative to patient volunteers and anonymous data.
PMID:42949820 | DOI:10.1002/mp.70742