Performance | NVIDIA Medical NIM Documentation

Accuracy Performance

To measure the accuracy of MAISI, synthetic images along with their annotated labels are generated. The VISTA-3D model is then used to evaluate the image quality. The table below shows the Dice score calculated between the MAISI-generated label and the VISTA-3D segmentation for specific anatomies. All tests are conducted with a spacing of 1mm.

Accuracy

Case	Body Region	Anatomy	Image Size	Dice Score
1	chest	aorta	256x256x256	0.885
2	chest	left lung upper lobe	256x256x128	0.777
3	abdomen	liver	256x256x256	0.909
4	abdomen	spleen	256x256x128	0.935

Speed Performance

The speed is measured for generating different body regions, anatomies, and image sizes. The table below shows the generation speed on different GPUs. All tests are conducted with a spacing of 1mm.

Case	Body Region	Anatomy	Image Size	Speed on A100 (seconds)	Speed on H100 (seconds)
1	chest	aorta	256x256x256	158	102
2	chest	left lung upper lobe	256x256x128	122	82
3	abdomen	liver	256x256x256	190	138
4	abdomen	spleen	256x256x128	123	82
5	chest	aorta	512x512x512	649	455
6	abdomen	spleen	512x512x512	651	456

Note: You may experience a slowdown when using Windows Subsystem for Linux (WSL).