Zero-shot evaluation of promptable foundation models for 3D CT segmentation

Caroline Magg, Clara I. Sánchez, Hoel Kervadec

Keywords: foundation models; segmentation; validation

TL;DR: We evaluated 11 promptable Foundation Models for bone segmentation in CT scans, using non-iterative 2D and 3D prompting strategies, and identified the Pareto-optimal models.

Abstract: Foundation Models (FMs) have revolutionized interactive segmentation for medical imaging. However, the increasing number of promptable FMs, along with evaluations varying in dataset, metrics, and compared models, makes direct comparison difficult and complicates the selection of the most suitable model for specific clinical tasks. In the context of bone segmentation in CT scans, we evaluated 11 promptable FMs using non-iterative 2D and 3D prompting strategies on both a private and public dataset. The models were categorized based on their prediction dimensionality (2D vs. 3D), and the Pareto-optimal models were identified.

Submission Number: 9