Medical Imaging

[AuliaRizky]

Is it possible to use in machine learning for medical imaging using MRI dataset?

[TarrySingh]

Yes, it is

[parinaya-007]

You can definitely use capsule net for medical imaging but then you will have to be careful about the architecture. You need to be aware of what type of entities the capsules represent and also the number of capsules in the final capsule layer are appropriate.
What exactly are you talking about in medical imaging? Like detection problems or generative problems?

[AuliaRizky]

I want to do segmentation on ischemic stroke using ISLES 2017 dataset. How do we decide those parameters such as number of capsules, the vector length, and else so that CapsNet could be implemented for 5D (or 4D) dataset. Of course conv3D will be used for architecture, but is it still possible to use basic architecture?

[parinaya-007]

@AuliaRizky In your case, I don't think the basic architecture will work. And as you said, conv3D should be applied for initial layers and then for primary capsule layer it's the matter of experiment in order to choose hyperparameters like number of capsules, vector length, etc.
What I can contribute in this is that you should keep in mind that each capsule in primary capsule layer represents probability and orientation information of a specific entity/object/part-of-object so you can take a look into your data and accordingly decide the number of such entities/object/part-of-object that would be required for the same.

[AuliaRizky]

Thanks a lot!
So from my understanding, in determining the number of hyperparameters I should test few (or lots) of combinations? Is there any method that I could use to optimize my experiment?

[parinaya-007]

@AuliaRizky Everytime when you apply a network to a new kind of dataset you need to run the model with various combinations of hyperparameters for the network. But I would like to suggest a research paper for you to understand the basics required for at least the convergence of capsule network on basis of hyperparameters. Following is the paper:
Empirical Study on convergence of Capsule Networks with various hyperparameters
http://people.cs.vt.edu/~bhuang/courses/opt18/projects/capsule.pdf

[AuliaRizky]

@parinaya-007 Can you explain to me how to changes information of pixels (scalar) to pose parameter (which is a vector) after applying convolution?
I have read "Transforming Auto-encoders" by Hinton that should explained it, but I am not fully understand. And after watching video from https://www.youtube.com/watch?v=pPN8d0E3900&t=722s , the creator said that we apply reshape to the feature maps. It still confuses me
And thanks for the paper!