Autonomous Scanning and Motion-Aware Segmentation for Robotic Thyroid Reconstruction

Ultrasound (US) imaging of thyroids, a non-invasive and real-time modality, has been one of the paramount techniques for nodule diagnosis. However, current procedures heavily rely on professional expertise. To mitigate this issue, we present an autonomous robotic system for real-time scanning and reconstruction of the thyroid from US images. We first develop an accurate PID controller scheme that integrates multimodal feedback from both image analysis and force sensors. This controller autonomously manages the thyroid scanning process using a robotic probe, without pre-planning the trajectory or any human intervention. Subsequently, we propose a real-time thyroid segmentation method termed TNSegNet, which exploits motion consistency across US sequences to improve segmentation coherence. Finally, we reconstruct and render the 3D thyroid according to the segmentation results. We perform extensive experiments to evaluate our proposed method and compare it with competing approaches on real-scanned thyroid datasets. Results demonstrate that our method achieves high-quality thyroid segmentation and reconstruction. Moreover, we showcase the application of our method to the scanning, segmentation, and reconstruction of thyroid nodules, where our method consistently delivers impressive performance. The code will be publicly available.