AI-DRIVEN RECONSTRUCTION AND AUTOMATED PHYSICS RIGGING OF DEFORMABLE COLON MODELS FROM COMPUTED TOMOGRAPHY FOR COLONOSCOPY REHEARSAL

Purpose: Current colonoscopy simulators typically use generic anatomy and limited soft-tissue collapse, constraining patient-specific rehearsal. We built an end-to-end pipeline that converts computed tomography (CT) colonography into a deformable three-dimensional (3D) colon model for interactive training and pre-procedure planning.

Methods/Materials: De-identified CT colonography volumes (n=14) were segmented via organ segmentation model to generate a binary lumen mask. A surface mesh was reconstructed (marching cubes) and smoothed/decimated to a simulation-ready resolution. A Python-based rigger extracted a centerline and local radius profile, partitioned the mesh into sequential rings, generated a skeletal chain, assigned vertex weights by nearest-surface mapping, and parameterized per-joint masses and hinge-spring constraints to approximate collapse and longitudinal stretch. The rigged mesh was then exported as a skinned model compatible with a real-time 3D simulation engine, where collisions with a colonoscope proxy drove deformation.

Results: The pipeline produced fully rigged, deformable, patient-specific colon models without manual 3D editing. In interactive testing, the model exhibited stable real-time deformation, including segmental luminal collapse under contact loading with recoil after unloading, enabling rehearsal of scope advancement and loop-management maneuvers along the subject-specific anatomy.

Conclusions: CT-derived, AI-segmented colon models can be automatically transformed into deformable simulators using a fully scripted physics rigging workflow. This scalable approach enables rapid creation of patient-specific rehearsal environments and provides a technical foundation for imaging-based 3D roadmaps to support navigation during colonoscopy. This underscores the intersection between radiology as a training overlay in other specialties.