SAM2CT Converts Routine Radiologist Annotations into 3D CT Lesion Segmentations

SAM2CT model converts radiologists' arrows and line annotations into 3D CT segmentations, achieving Dice coefficients of 0.649 (arrow) and 0.757 (line). On clinical PACS data (n=60), 87% of generated segmentations were clinically acceptable. Zero-shot ED findings showed DSC 0.61…

• Retrospective study using public benchmarks and clinical PACS data (n=60), with a zero-shot emergency department cohort (n=130) for abscess and gallstone findings.
• SAM2CT extends SAM2 with a prompt encoder for arrows/lines and a memory-conditioned memory module, improving slice-to-slice propagation accuracy.
• Single-center, retrospective design with limited external validation; zero-shot performance was evaluated only on selected ED findings, and sample size for clinical PACS data was small.

RadPigeon summaries are original and for information only. They are not clinical advice.