To develop a framework that enables efficient 3D segmentation from minimal 2D prompts, specifically addressing the limitations of extensive manual annotations and task-specific retraining in current medical imaging methods.
Key Findings:
PAM outperformed MedSAM and SegVol, improving average DSC by 19.3% (P < 0.001).
Stable performance under variations in prompts (P ≥ 0.5985) and propagation settings (P ≥ 0.6131).
Achieved faster inference (P < 0.001) and reduced user interaction time by 63.6%.
Gains were strongest for irregular objects, with improvements negatively correlated with object regularity (r < -0.1249).
Interpretation:
PAM provides a robust and generalizable tool for automated clinical imaging, significantly reducing reliance on manual annotation and task-specific training, which can enhance workflow efficiency in clinical settings.
Limitations:
Performance may vary with different imaging modalities not tested, potentially affecting generalizability.
Further validation needed on a wider range of anatomical structures to ensure robustness.
Conclusion:
PAM effectively addresses the challenges of 3D segmentation in medical imaging, offering a promising solution for clinical applications.
