search

naturomics / CapsNet-Tensorflow

A Tensorflow implementation of CapsNet(Capsules Net) in paper Dynamic Routing Between Capsules
Apache License 2.0
3.8k stars 1.17k forks source link

Medical Imaging #75

Open AuliaRizky opened 5 years ago

AuliaRizky commented 5 years ago

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

TarrySingh commented 5 years ago

Yes, it is

parinaya-007 commented 5 years ago

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 commented 5 years ago

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 commented 5 years ago

@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 commented 5 years ago

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 commented 5 years ago

@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 commented 5 years ago

@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!