Development of 3D-printed Individualized Vascular Phantoms for Artificial Intelligence (AI) Enabled Interventional Device Testing

We developed vascular phantoms mapped from human subjects to test AI-enabled ultrasound guided vascular cannulation. Translational device prototyping necessitates anatomically accurate models. Commercial phantoms fail to address anatomic variability. Uniformity leads to optimistic AI model and operator performance.  Individualized 3D-printed vascular phantoms yield anatomically correct models optimized for AI-device testing.

Materials and Methods: After IRB approval, thin section (≤ 1mm) CT scans with inguinal arterial and venous enhancement were identified. Vessel lumens, skin surface, and bones were segmented. The external mold and bone elements were printed in polyethylene terephthalate glycol-modified (PETG) and 0.6 mm thick hollow vessel lumen scaffolds with water-soluble butenediol vinyl alcohol copolymer (BVOH).  The mold interior with skin contour and vessel scaffolds were coated in silicone (DragonSkin 20).  Vessels were mounted and the mold box was filled with gelatin/agar, graphite scatterers and 1-propanol.  After cooling, water flushing the vessel dissolved BVOH, leaving embedded patent silicone vessels.  Water was pumped intravascularly during AI-GUIDE {1} testing and operators were blinded to B-mode images. AI-GUIDE segments vessels on live ultrasound B-mode images with real-time AI, guiding an operator to position the device to an optimal location followed by integrated surgical robot vascular cannulation. Successful cannulation was assessed by fluid return. The AI algorithm was iteratively re-trained, but not on the phantom being tested.

Results:

With iterative prototyping, 2 custom vascular phantoms were amenable for testing. 6 operators with no (3), 1-5 years’ (2), and 5+ years’ (1) experience performed 27 injections; 81% were successful. 2 failures were due to needle insertion motor malfunction and 3 consecutive attempts were later recognized to be caused by a damaged needle. Median time to cannulation was 48 seconds (range 19-275 seconds).

Conclusion:

Custom phantoms allow device testing on diverse human subject anatomy, critical for AI-processes prone to over-fitting. We successfully pilot phantom fabrication and AI-GUIDE testing showing that novice users can rapidly cannulate femoral vessels with high accuracy. Ongoing work seeks to reduce artifact generation during the phantom fabrication process, improve the generalizability of the AI-algorithm, and perfect hardware performance. This sets the stage for future noninvasive “phantom clinical trials” of AI-enabled interventional devices.