Performance

With an average GPU GeForce GTX 1050 Ti of 768 cores/ 4GB buffer, 7Gbps memory speed, Intel® Core™ i7-8700 CPU @ 3.20GHz × 12, and a system memory of 16 GB, Brainchop shows in general fast response while processing 256x256x256 raw Nifty image.

Sub-volumes Version:

The inference iterated through the volume 341 patches, each of shape 38x38x38. The pipeline takes around 35~ 40 seconds to complete the whole task and is tested with multiple browsers that support WebGL-2.

Adding 200 overlapping cubes for better accuracy can take around 120 seconds.

Full-volume Version:

For the MeshNet model of 5 filters, the pipeline takes around 30 seconds to complete the whole task and is tested with multiple browsers that support WebGL-2.

The MeshNet model of 11 filters takes around 40 Seconds to process the whole volume.

Model Accuracy

While MeshNet Model has fewer parameters than the classical segmentation model U-Net, it can achieve a competitive DICE score. For more information, please refer to MeshNet models segmentation accuracy and code here.