Fig. 4: Quantitative analysis of PAM across various datasets.

a Radar chart comparisons of Dice Similarity Coefficient (DSC) among MedSAM, SegVol, PAM-2DBox, and PAM-2DMask across 44 datasets (D01–D44), with DSC values ranging from 0.0 to 1.0, moving from the center outward. b Performance comparison on ten external datasets across four metrics: Sensitivity, Normalized Surface Dice (NSD), 95th percentile Hausdorff Distance (HD95), and Average Surface Distance (ASD). Both HD95 and ASD are normalized by their respective maximum values within each dataset. c Comparison of model performance under the same 3D box prompt. The top shows a radar chart comparing PAM-3DBox with baseline prompt-based models using identical 3D box input. The fully-supervised model nnUNet is included as a task-specific performance upper bound. The bottom bar chart illustrates the performance change of PAM-3DBox compared to PAM-2DBox across datasets. d Comparison of inference times (seconds). The left side showing a box plot of inference time distribution across 44 datasets, and the right visualizes a comparative analysis of the inference times for each model across these datasets. e Comparison of manual prompt times (seconds). A box plot depicts the distribution of interactive prompt times for three distinct prompt types. f Stability analysis with respect to initialization slice deviation. Box plots show the DSC distribution for PAM-2DBox (blue) and PAM-2DMask (orange) across deviation levels: 0% (no deviation), ±5%, ±10%, ±15%, and ±20%. g Consistency analysis with respect to propagation slice thickness. Box plots show the DSC distribution across thicknesses of 10 mm, 20 mm, 30 mm, and 40 mm. The visualization was generated using ITK-SNAP (version 3.8.0).