Add texture based feature extraction (radiomics) about gandlf HOT 5 CLOSED

Hi @sarthakpati ,

Sorry for the late reply. While hand engineered texture features are not commonly used for segmentation tasks, but can be passed on as an additional input to the CNN for classification tasks.
Thanks i'll check it out.

Ahh I have used CaPTk previously for automated segmentations didn't it had radiomics too. Thanks!

from gandlf.

Hi @umgpy,

Thanks for your interest in GaNDLF!

Is there value in adding engineered features for a DL algorithm? The reason why I ask this is because the entire idea of convolutional filters is to allow a network to judge what the "best" combination of filters is going to be for a particular problem statement. What this means is that the same network architecture can end up dramatically different final filters for brain tumor segmentation and brain extraction (you can read more about this on any of the Grad-Cam or related papers). Taggin @Geeks-Sid to provide more elaboration and details.

FYI, we have our own radiomics package, CaPTk, which is also IBSI compliant and pretty good, too. 😉

Cheers,
Sarthak

from gandlf.

but can be passed on as an additional input to the CNN for classification tasks

Could be interesting to explore. I don't think anyone in our side currently has this bandwidth, but if you can put in a PR for this and generate some results, it would make for a great contribution!

from gandlf.

Thanks I'll try to implement a working version!

from gandlf.

Feel free to open a PR when you are able to add this. Closing for now.

from gandlf.