search

neuronflow / BraTS-Toolkit

Code to preprocess, segment, and fuse glioma MRI scans based on the BraTS Toolkit manuscript.
https://www.frontiersin.org/articles/10.3389/fnins.2020.00125/full
GNU Affero General Public License v3.0
70 stars 11 forks source link

Minimal segmentation requirements #45

Open BegoneGIT opened 7 months ago

BegoneGIT commented 7 months ago

Hi! I'd like to ask what are reccomended specs to run segmentor on? I've i5-9300HF, 8GB GeForce GTX 1650, so not an amazing spec.

I ask this since isen-20 hangs after ~20 minutes (docker image CPU and RAM usage goes to 0, but virtual machine itself still hogs on those resorces somehow). To free up the memory I have to quit Docker altogether. I tried deleting and recreating this docker image, but it doesn't seem to help. Last thing printed in console is Executing: docker run --rm --gpus device=0 -v \BraTS-Toolkit\brats_toolkit\tmpps3eqv8s:/app/data/ brats/isen-20 python runner.py.

I think it's memory issue because sanet0-20 returned CUDA out of memory error.

I was able to run examples and they did return proper segmentation masks in nii.gz.

Thinking about possible solution I came up with two things. For first maybe I could 'reduce' the data somehow so models have more memory to work on. Second is, maybe there are smaller (or pruned) models I could use.

neuronflow commented 7 months ago

The minimum hardware requirements will largely depend on the algorithm, operating system, etc., e.g., you have a VM in your setup, which creates additional overhead.

There is no general answer to this question. Unfortunately, we don't have a good overview of algorithm-specific hardware requirements. A PR featuring such an overview of the documentation would be welcome.

Hopefully, we will find time to publish a selection of glioma segmentation algorithms that don't require docker and in general, are very light-weight soon [we have an upcoming publication with already trained networks].