Table 10.

Training and inference time of different model scales. Training iteration time reports the time for processing a single 96×96×96 patch. For the testing cases, the latencies are reported on MNI space size of 172 × 220 × 156 using GPU RTX 3090Ti, regular inference pipeline without mixed precision, and no torchscript and TensorRT conversions. Note the inference patch size is 96 × 96 × 96, sliding window overlap has a significant impact on the inference time because the increase of overlap percentage will result in exponentially increased patches.

Model	UNesT-S	UNesT-B	UNesT-L
Training
Iteration (s)	0.29	0.46	0.82
Testing
overlap = 0.3 (s)	0.84	0.98	1.23
overlap = 0.5 (s)	2.30	2.34	2.97
overlap = 0.7 (s)	6.76	6.99	7.85