Intelligent control of robotic X-ray devices using a language-promptable digital twin.

Abstract

Purpose: Natural language offers a convenient, flexible interface for controlling robotic C-arm X-ray systems, making advanced functionality and controls easily accessible.Please confirm if the author names are presented accurately and in the correct sequence (given name, middle name/initial, family name). Author 1 Given name: [Benjamin D.] Last name [Killeen]. Also, kindly confirm the details in the metadata are correct. However, enabling language interfaces requires specialized artificial intelligence (AI) models that interpret X-ray images to create a semantic representation for language-based reasoning. The fixed outputs of such AI models fundamentally limits the functionality of language controls that users may access. Incorporating flexible and language-aligned AI models that can be prompted through language control facilitates more flexible interfaces for a much wider variety of tasks and procedures.

Methods: Using a language-aligned foundation model for X-ray image segmentation, our system continually updates a patient digital twin based on sparse reconstructions of desired anatomical structures. This allows for multiple autonomous capabilities, including visualization, patient-specific viewfinding, and automatic collimation from novel viewpoints, enabling complex language control commands like "Focus in on the lower lumbar vertebrae."

Results: In a cadaver study, multiple users were able to visualize, localize, and collimate around structures across the torso region using only verbal commands to control a robotic X-ray system, with 84% end-to-end success. In post hoc analysis of randomly oriented images, our patient digital twin was able to localize 35 commonly requested structures from a given image to within $51.68\pm 30.84$  mm, which enables localization and isolation of the object from arbitrary orientations.

Conclusion: Overall, we show how intelligent robotic X-ray systems can incorporate physicians' expressed intent directly. Existing foundation models for intra-operative X-ray image analysis exhibit certain failure modes. Nevertheless, our results suggest that as these models become more capable, they can facilitate highly flexible, intelligent robotic C-arms.