Anatomy-Aware Robotic Ultrasound Path Planning With Diffusion-Based View Completion

Abstract

Autonomous ultrasound (US) probe initial localization is challenging when external cameras are obstructed or provide only partial patient views, as this limits the utility of anatomical priors. To overcome this, we introduce a novel robotic US probe localization framework featuring two key innovations. First, we employ a human anatomical model comprising bones and arteries to localize target anatomy, enabling the autonomous generation of clinically relevant scan paths. Second, we employ a diffusion-based view completion module to create a complete and anatomically correct patient view from a partial image. This module features a LoRA fine-tuned Stable Diffusion network guided by a pose-conditioned ControlNet for intelligent outpainting. The completed view enhances the accuracy of the human anatomical model fitting, thereby improving the robustness and precision of localization. Our approach was validated through experiments on datasets and in-vivo volunteers, targeting both vascular and skeletal structures. Our method achieved a 100% success rate, with localization errors of 7.73 mm for the carotid artery and 9.20 mm for the lumbar spine. In comparison, the baseline method without view completion exhibited a lower success rate and higher errors of 12.41 mm and 14.69 mm, respectively. These findings indicate that combining anatomy-aware path planning with diffusion-based view completion facilitates reliable autonomous localization in realistic, partially occluded environments.