Reproduce “UltraBones100k: A reliable automated labeling method and large-scale dataset for ultrasound-based bone surface extraction” on OpenResearch

About this paper

Luohong Wu, Nicola A. Cavalcanti, Matthias Seibold, Giuseppe Loggia, Lisa Reissner, Jonas Hein, Silvan Beeler, Arnd Viehöfer and 3 more

UltraBones100k introduces a large-scale public dataset of 100,000 ex-vivo ultrasound images with automatically generated, high-quality bone annotations derived from an optimized pipeline. A deep learning model trained on this dataset achieved superior bone segmentation performance, particularly in challenging low-intensity regions, outperforming expert manual labeling and demonstrating qualitative generalizability to new ex-vivo and in-vivo anatomies.

Key insights

• The automated labeling method, enhanced by per-frame optimization and physics-based refinement, produces more complete and accurate bone annotations than human experts, particularly in challenging low-intensity and anechoic regions.
• Training a deep learning model on a large-scale, precisely annotated dataset like UltraBones100k significantly improves bone segmentation accuracy and robustness, demonstrating consistent superiority over traditional manual labeling.
• Models trained exclusively on ex-vivo ultrasound data can exhibit promising qualitative generalizability to new ex-vivo anatomies and, critically, to in-vivo human data, suggesting strong potential for clinical translation.

Summary by alphaXiv.