. Author manuscript; available in PMC: 2025 Aug 26.

Published in final edited form as: Med Image Comput Comput Assist Interv. 2024 Oct 23;15012:542–552. doi: 10.1007/978-3-031-72390-2_51

Table 3.

Comparison of the computational efficiency with respect to the encoder and decoder.

Dim	Method	Resolution	Encoder			Decoder			Acceleration
Dim	Method	Resolution	Time (ms)↓	FLOPs (G)↓	Memory (Gb)↓	Time (ms)↓	FLOPs (G)↓	Memory (Gb)↓	To 2D ↑	To 3D ↑
2D	`SAM` [16]	1024 × 1024	3980	369.0	7.87	239	3.0	5.57	1.00×	/
	`MobileSAM` [32]	1024 × 1024	584	36.7	5.48	233	3.0	5.27	5.16×	/
	`TinySAM` [24]	1024 × 1024	609	36.7	5.48	246	3.0	5.27	4.93×	/
	`MedSAM` [18]	1024 × 1024	3983	369.0	7.87	241	2.9	5.57	1.00×	/
	`SAM-Med2D` [4]	256 × 256	1063	32.0	6.32	216	0.21	5.55	3.30×	/
3D	`SAM-Med3D` [27]	128 × 128 × 128	70	89.5	6.58	20	2.8	5.53	60.27×	1.00×
	`FastSAM3D`	128 × 128 × 128	3	21.9	0.78	5	2.8	0.71	527.38×	8.75 ×

We report the time (ms), FLOPs (G), and memory (Gb), alongside acceleration factors relative to 2D SAM [16] and 3D SAM-Med3D [27]. For 2D SAMs, we compute the time to process all the slices within volumetric data. The best results are highlighted in bold if statistically different from the second best result (p < 0.01).