search

IAmSuyogJadhav / 3d-mri-brain-tumor-segmentation-using-autoencoder-regularization

Keras implementation of the paper "3D MRI brain tumor segmentation using autoencoder regularization" by Myronenko A. (https://arxiv.org/abs/1810.11654).
MIT License
369 stars 113 forks source link

decoder blue boxes #15

Open alalbiol opened 5 years ago

alalbiol commented 5 years ago

Hi again Checking your model implementation I realized that you decoder blue box outputs 32 channels and then you apply a final conv layer into the 3 output channels

I think that the blue box is the one that maps from 32 to 3 output channels since no extra layer is drawn in the paper.

The paper says: "The end of the decoder has the same spatial size as the original image, and the number of features equal to the initial input feature size, followed by 1x1x1 convolution into 3 channels and a sigmoid function."

I think that "the end of the decoder" here means the output of the last green block and the 1x1x1 convolution that follows is the blue box

I have also found this implementation that agrees with my interpretation: https://github.com/athon2/BraTS2018_NvNet/blob/master/nvnet.py

What do you think?

IAmSuyogJadhav commented 5 years ago

@alalbiol, Sorry for my delayed response. I was busy in an academic event for the duration of the last 6 days, so could not attend to any of the issues during that time.

Thank you for pointing it out. I had a look at my code and realized that not only am I using the wrong blue block, but also the green decoder blocks in my implementation use 1x1x1 convolutions instead of 3x3x3. I will need to change the decoder blue and green blocks and make sure the output shapes match up. I have a couple of things at hand right now. So, might take some time for me to start working on this issue. Added to To-Do anyways. Thank you for your contribution.

IAmSuyogJadhav commented 5 years ago

Meanwhile, if anyone fixes this or any other issue, please submit a PR. This will be helpful for everyone.