Training of CAE

[shubhamjain0594]

I am trying to implement your paper in Torch and wanted some help in it.

Firstly I want to know how are you exactly training the CAE, greedy layerwise or end to end.

[ehosseiniasl]

CAE are trained greedy layer-wise. The encoding layers are stacked and frozen during fine-tuning

[shubhamjain0594]

Thanks for the reply.

Also you mention about pre-training, initial training and fine-tuning. I understand the part about pre-training on CADDementia and fine-tuning on ADNI but what do you mean by initial training of lower convolutional layers and how do you do it.

[ehosseiniasl]

there is no initial training! only pertraining and then fine-tuning

[shubhamjain0594]

One last question, you mention about using pooling and conv in a CAE, so is your architecture Input -> Full Conv -> MaxPool -> MaxUnpool -> Validconv -> Output

Because you mention for input size of 200*150*150 you get feature maps of size 102*76*76, 52*40*40 and 28*22*22.

[shubhamjain0594]

Also, can you add the list of ADNI files you have used? ADNI Data has 3000+ MRIs so it will be great if you can share the list of files you have used.

[ehosseiniasl]

@shubhamjain0594 In the current implementation of CAE, no pooling layer is used. to get hidden activation of a CAE to train the subsequent CAE, max_pooling is used. you can check the do_pretraining_cae() in main.py. Each cae has full_conv --> valid_conv structure.

[ehosseiniasl]

@shubhamjain0594 list of ADNI subject IDs are added!

[christiantinauer]

@ehosseiniasl Could you please also add the image IDs? Because the infos are ambiguous for some of the patients.

[mjachowdhury]

@ehosseiniasl Hi, I Would like to know that the subject ID, e.g., 002_S_0619, has 36 images for AD on different weighting, slice thickness, and acquisition. Did you select all the images for that subject ID or only one image from that subject ID as input for the model? I look forward to hearing from you. Thank you