ANTsPyNet's super resolution models

[carlouic]

Hi everyone,

Thanks so much for your effort in providing us such a great tool. I have a few questions regarding the models for super-resolution. I saw that you have implemented several architectures, and that you have also uploaded the pre-trained weights for the Deep Back Projection (DBPN) Network.

• Are you going to upload also the pre-trained weights for other models? (SRCNN, ResNet SR...)
• Are you going to upload also the training scripts for these models?
• Could I know the datasets on which the DBPN has been trained?

Thanks for your time and your help.

[ntustison]

Are you going to upload also the pre-trained weights for other models? (SRCNN, ResNet SR...)

Pre-trained weights don't exist for the other models. Preliminary exploration led to the conclusion that the DBPN architecture performed the best for our task but none of that preliminary work was saved.

Are you going to upload also the training scripts for these models?

I've thought quite a bit on how to include training scripts as a formal part of ANTsXNet but just haven't found a satisfactory solution. For some of my other papers, such as this one, I've posted the scripts to the GitHub repo associated with the paper. @stnava is in the process of putting together the super resolution paper and I suspect he might do something similar.

Could I know the datasets on which the DBPN has been trained?

I believe it was the Human Connectome Project but I'll let @stnava confirm whether or not this is the case.

[carlouic]

Thanks @ntustison for the kind reply. Can I just ask if you've followed the training strategy that they have employed in the DBPN paper and github? @stnava, is it correct that the data you have used to train the DBPN was the HCP? If so, can I ask how you have preprocessed them before feeding it to the model? Thanks again for your help.

[carlouic]

Hello again. If you could provide a little bit of details about the dataset population, such as number of images, demographics of the subjects you chose to train the network on, train-test-validation split percentages... that would be highly appreciated.

I read an abstract of yours, from 2020, where you mention you trained the architecture on a total of 120 subjects: Deep volumetric super-resolution improves the detection of amyloid-related cross-sectional group differences in MCI. Since there you also mention you trained the model on the ADNI dataset, I guess the model available through antspynet pretrained networks is another one, right?

[ntustison]

Brian's pretty busy so I don't know when he'll be able to respond. However, I can answer

Can I just ask if you've followed the training strategy that they have employed in the DBPN paper and github?

I didn't do any of the training. I merely ported the architecture and Brian did all the training.

[ncullen93]

Closing due to inactivity, but improved documentation of pre-trained models is definitely on the roadmap.