AHNet as 3D feature extractor

[joaolcguerreiro]

Hi,

is it possible to use AHNet as a pre-trained 3D feature extractor?

In such case, what should be the transforms applied to the input tensors before feeding them to the pre-trained network? Should pixel values be rescaled? What were the training normalization statistics?

Also, which layers should be considered for feature extraction? Every layer except the last?

Thank you! :)

[joaolcguerreiro]

I have asked this question in Model Zoo, since it may be more adequate :)

[KumoLiu]

Hi @joaolcguerreiro, thanks for your interest here!

is it possible to use AHNet as a pre-trained 3D feature extractor?

IMO any CNN can be used as a feature extractor.

In such case, what should be the transforms applied to the input tensors before feeding them to the pre-trained network? Should pixel values be rescaled? What were the training normalization statistics?

For how to preprocess your data, I think it depends on your data instead of the network. For example, you can apply a Z-score normalization based on the mean and standard deviation of all collected intensity values for CT and for MRI or other image modalities simple Z-score normalization can be applied to the patient individually.

Also, which layers should be considered for feature extraction? Every layer except the last?

IMO, yes. Each layer can be used as a feature extractor, except for the last layer which outputs logits.

Hope it helps, thanks!