. 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 4.

Ablation study for the contribution of 3D sparse attention (‘Sparse Attn.’) and flash attention (‘Flash Attn.’) to the performance and efficiency of FastSAM3D.

Sparse Attn.	Flash Attn.	AMOS [13]				TotalSegmentator [28]				BraTS [20]				Encoder
Sparse Attn.	Flash Attn.	1pt	3pt	5pt	10pt	1pt	3pt	5pt	10pt	1pt	3pt	5pt	10pt	Time	FLOPs	Memory
✗	✗	0.282	0.375	0.403	0.436	0.243	0.371	0.442	0.516	0.335	0.404	0.422	0.444	10	23.1	1.16
✗	✓	0.276	0.366	0.398	0.432	0.247	0.374	0.438	0.516	0.331	0.402	0.421	0.445	6	21.9	1.15
✓	✗	0.277	0.370	0.402	0.433	0.255	0.381	0.450	0.520	0.328	0.403	0.422	0.445	9	23.1	0.79
✓	✓	0.273	0.368	0.402	0.437	0.250	0.378	0.445	0.519	0.333	0.401	0.421	0.445	3	21.9	0.78

Best scores are highlighted in bold, if statistically different from the second best result (p < 0.01). 3D sparse attention and flash attention contribute to substantial improvements in time and memory requirements without statistically significant performance decline.