SAM Zero-Shot detection of vertebral surfaces from freehand 3D ultrasound

SAM ZS detection of vertebral surfaces from FH 3D US

Abstract

Freehand 3D ultrasound imaging has emerged as a promising modality for spine imaging due to its non-invasive nature and cost-effectiveness. Accurate extraction of the 3D bone surface from freehand ultrasound image sequences is crucial for comprehensive interpretation of spinal structures, geometric spine deformity analysis, and for guiding ultrasound-assisted spine interventions. However, the vertical bone surface adjacent to the spinous process is usually invisible in ultrasound scans. In this preliminary study, a novel method for tissue-bone separation utilizing landmark-guided SAM is proposed to achieve accurate 3D bone surface reconstruction from low-quality ultrasound image sequences. The resulting bone surface, represented as a 3D mesh, effectively reproduces a smooth outline of the vertebra, encompassing the left lamina, spinous process, and right lamina. To evaluate the performance of the proposed method, we collected three freehand ultrasound sequences from three participants. The acoustic shadow masks beneath the extracted bone surface were evaluated against manually labeled masks, achieving a mean IoU above 0.92. Our evaluation on one human subject’s data reveals a mean surface distance of the extracted lumbar vertebral surfaces of 0.61±0.12 mm.